JP6828065B2 - Systems and methods for recursive cortical networks - Google Patents

Description

[0001] The present invention relates generally to the field of artificial intelligence, and more specifically to new and useful systems and methods for recursive cortical networks in the field of artificial intelligence.

Despite advances in computer vision, image processing, and machine learning, there are still challenges in recognizing visual objects that computers are inferior to human capabilities. Recognizing an object from an image not only requires recognizing the image in the scene, but also requires recognizing the object at various positions, with different settings, and with slight changes. For example, in order to recognize a chair, it is necessary to understand the innate nature of the chair as a chair. This is an easy task for humans. Computers struggle to deal with different types of chairs and situations where chairs can exist. This problem is even more difficult given the problem of detecting multiple objects in the scene. There are models for object recognition such as convolutional neural networks, HMAX models, slow feature analysis (SFA), and hierarchical time memory (HTM), but these approaches achieve results close to ideal recognition performance. Can't. Object detection is more broadly a matter of pattern detection. Pattern detection is a problem in other areas and media other than image processing, such as speech recognition, natural language processing, and other areas. Furthermore, the reverse of pattern recognition is generation. Pattern generation has similar problems, and existing methods cannot produce satisfactory results as well. Therefore, in the field of artificial intelligence, there is a need to create new useful systems and methods with improved object recognition (or "inference") and generation. The present invention is referred to as a recursive cortical network and provides such a system and method.

1A and 1B are schematic views of a system of preferred embodiments. FIG. 2 is a schematic representation of a generalized representation of the system of preferred embodiments. FIG. 3 is a schematic diagram of a subnetwork of a preferred embodiment. FIG. 4 is a schematic diagram of a subnet for generating a pattern using the laterally constrained nodes of the preferred embodiment. FIG. 5 is a schematic diagram of a subnet for generating a pattern using the external constraint nodes of the preferred embodiment. FIG. 6 is a schematic diagram of a subnet for generating a pattern using the time constrained nodes of the preferred embodiment. FIG. 7 is a schematic diagram of a subnet for inferring a pattern of a preferred embodiment. FIG. 8 is a schematic diagram of a modified example of a network having two subnetworks sharing a child feature node of a preferred embodiment. FIG. 9 is an exemplary schematic of a network variant having a plurality of subnetworks divided between two layers, some of which share child feature nodes. FIG. 10 is a schematic diagram of a method of creating a network of preferred embodiments. FIG. 11 is a diagram showing an exemplary implementation example of a method for creating a network. FIG. 12 is a schematic diagram of a method of generating a pattern of a preferred embodiment. FIG. 13 is a schematic diagram of a method for inferring a pattern from a network of preferred embodiments. FIG. 14 is a schematic diagram of a network with various constraint nodes.

The following description of preferred embodiments of the invention is not intended to limit the invention to these preferred embodiments, but rather allows one of ordinary skill in the art to manufacture and use the invention. Intended to be.

l. System of Recursive Cortical Network [0018] As shown in FIGS. 1A and 1B, a system of preferred embodiments comprises a recursive cortical network 10 consisting of a plurality of subnetworks 100. The subnetwork preferably includes at least one parent feature node 110, a pool node 120, and a parent-specific child feature node 130 (PSCF node for short). It has a constraint node 140. The system functions to improve the invariance, selectivity, and sharing of information in the network. A network is, in a sense, a network of distributed processing elements that perform addition, multiplication, exponentiation, or other functions on incoming messages / signals. Patterns can be inferred and / or generated by propagating node activations over the network. Networks that can be modeled as neural networks or Bayesian networks can be enabled and implemented through various implementation examples. In the first implementation, the system is implemented as a network of electronically coupled feature node elements. A feature node element can be a logic gate arranged or configured within a processor to perform a particular function. In the second implementation, the system is implemented as a network model programmed or configured to run on a processor. The network model is preferably electronically stored software that encodes the operation and communication between the nodes of the network. The network 10 can be used for a wide variety of applications, including images, videos, sounds, natural language texts, analytical data, widely distributed sensor data, or other suitable forms of data. Data types can be used. Further, the network 10 can be configured for different operating modes, including a first mode of operation: generation mode and a second mode: inference mode. As shown in FIG. 1A, the network 10 is preferably a hierarchically organized network of interconnected sub-networks having various parent-child relationships. The network can also be a single layer or a single subnetworks of a collection of subnetworks as shown in FIG. 1B. The resulting morphology of the network 10 described herein is in part with a recursive cortical network (RCN) by a unique form of neural network, i.e., a repeating subnetwork pattern used to implement a cortical network. Can be called.

The hierarchical network 10 of the preferred embodiment functions to build subnetworks within the various layers. As shown in FIG. 2, various instances and instantiations of subnetwork 100 are recursively constructed, connected, and used in the hierarchy of network 10. The architecture of the hierarchical network 10 can be constructed algorithmically, or at least through partial user selection and configuration. The hierarchical network 10 can be described as alternating layers of feature nodes and pool nodes in a neural network. Subnetworks have feature input nodes and feature output nodes, which are used to bridge or connect subnetworks. As shown in FIG. 2, feature nodes can be constrained to various immutable patterns by using constraint nodes that bridge constraints between subnetworks that differ spatially or temporally across the pool. Each node in the hierarchical network preferably has a parent node connection and a child node connection. In general, the parent node connection is preferably the input in generation and the output in inference. Conversely, child node connections are outputs in generation and inputs in inference. In a variant of the single-layer (or non-layered) sub-network 100, the sub-network 100 is arranged as siblings. Subnetworks 100 as described below can interact through various forms of constraint nodes.

The subnet 100 can be set up in a variety of different configurations within the network. Many configurations are determined by constraint nodes that define node selection within, between, or even between subnetworks. In addition, subnetworks can be configured to have individual or shared child features. Subnetworks can be arranged more hierarchically. In other words, the first subnetworks can be the parent of the second subnetworks. Similarly, the second subnetworks can be the parent of the third subnetworks. Subnetwork layers are preferably connected via shared parent and child feature nodes. Preferably, the child feature node of the top layer subnetwork is the parent feature node of the lower subnetwork. Conversely, the parent feature node of the subnetwork 100 can participate as a child feature node of the higher subnetwork 100. The parent feature node of the top-level subnetworks is preferably the input to the system. The child feature of the lowest / lowest subnetwork is preferably the output of the system. By connecting a plurality of subnetworks 100, multi-parent interaction can be introduced at some nodes in the network. These interactions can be modeled using different probabilistic models within the node.

When the sub-networks 100 are connected hierarchically, they function to promote compact and compressed representation through the reuse of the sub-networks. A parent feature node of one subnetwork can participate as a child feature node in multiple parent subnetworks. A similar advantage is that the invariant representation of the child subnetworks can be reused in multiple parent subnetworks 100. An example of where this is applicable is when network 10 represents a visual object. The lower subnetworks 100 can correspond to parts of the object, and the higher subnetworks (ie, upper subnetworks) can represent how those parts come together to form the object. For example, lower level subnetworks can accommodate representations of body parts in cow images. Each body part is expressed immutably and is tolerant of position changes such as translation, scale change, and distortion. Higher-level subnetworks specify how each body part gathers to represent a cow. Some of the lower level body parts of cows can be reused at higher levels to represent goats. For example, the paws of both of these animals move in the same way, so these parts can potentially be reused. This means that the invariant expressions learned for the cow's paws are automatically reused for the goat's expression.

The system can be used for inference or generation. Inference can include other uses, including pattern detection, classification, prediction, system control, decision making, and inferring information from data. The generation may include still images, video graphics, audio media, text content generation, action or response selection, or any suitable medium that is synthesized based on higher level inputs. In a preferred implementation, the network 10 can be selectively used for inference or generation, and in some variants, both modes of operation can be realized simultaneously. When the network 10 is used for inference applications, the operation of the network 10 preferably begins with sample data reduced, transformed, or extracted into data features. Data features are preferably attribute specifications and their values. A feature vector is preferably a set of features for an instance of a data sample. For example, when applying a network to inference of image data, the image is subdivided into a plurality of image blocks, and a pixel pattern in the plurality of blocks is used as a feature. The input data features are preferably transmitted, supplied, or directed to the corresponding child feature nodes of the network 10. In other words, the data features are directed to the bottom child feature nodes of the subnet 100. In the inference operation, the node operates informed and preferably propagates the node selection / processing through the hierarchy of the network 10 until the output is obtained from the parent feature of the top-level subnet 100. There is also a combination of information propagation to higher levels (upper parent layers) and lower levels (last child features) in the hierarchy. By expanding the hierarchy downward during inference, the network can improve accuracy by comparing the inference conclusions for the input data with the input element features using the imagination / generation function. In inference, the output is preferably inference information. For example, if inference is used to detect objects in an image, the output can be the identification of the detected objects. When the network 10 is used for generation purposes, the operation of the network 10 preferably begins with a general generation request directed to, supplied, or delivered to the parent feature node of the top-level subnet 100. In the generation operation, the nodes preferably operate informed and propagate the node selection / processing down the hierarchy of network 10 until the output is obtained from the child feature nodes of the bottom layer subnetwork 100. More specifically, the top-level subnet 100 produces multiple samples at the same time. The output samples of these top layer subnetworks 100 determine which lower layer subnetworks 100 are active. Next, samples are generated simultaneously from these lower layer subnetworks 100. This output further determines the lower layer active subnet 100. This pattern continues through each layer of network 10 until a sample is finally generated from the bottom layer of subnet 100. In generation, the output is preferably a simulated output. For example, if network 10 is used to generate an image and the input is the name of an object, the output is preferably an image representing that object name. More preferably, the generation and inference include a hybrid or mixed input approach. The generation behavior further relies on inputs to child feature nodes that are partially completed from the previous instance, noisy, distorted, or otherwise serve as a general guide to the generation process. In one variant, the generation preferably involves the system acquiring an image input for half of the image. This half of the image input is input to the child features. Network 10 is then prompted to create possibilities for the other half. In some variants, the network 10 may include subnetworks connected so that the connection skips one layer. The output of the sub-network 100 that skips the layer will preferably be involved in the selection of the active sub-network 100 when the other sub-network is fed to the corresponding layer.

As shown in FIG. 3, the sub-network 100 functions to provide a node selection operation between the parent feature and the child feature. The sub-network 100 is a basic component of the network 10. Subnetwork 100, in the case of generation, preferably has a higher level such that a lower level feature activity (eg, visual features of an image) is determined from a higher level feature activity (such as an object name). Features are mapped or networked to a set of lower level features. In the case of inference, the subnetwork is preferably at a lower level, such that the higher level feature activity (eg object name) is determined by the lower level feature activity (eg visual feature of the image). It is mapped or networked from features to higher level features. The general architecture of the subnet 100 preferably includes a single top-level node, which is the parent feature node 110. The parent feature node 110 (PF1) preferably includes connections to at least two pool nodes 120 (P1 and P2). Each pool node 120 preferably includes connections to a plurality of PSCF nodes 130 (X1, X2, X3, X4, X5, X6). Constraint nodes 140 (C1, C2, C3) may also be in the subnet 100. The constraint node 140 is preferably connected to another PSCF node 130. Constraint node 140 defines restrictions, rules, and restrictions between at least two PSCF nodes 130. The PSCF node 130 is preferably connected to the child feature nodes 150 (CF1, CF2, CF3, CR4, CF5, CF6). Instances of subnetworks 100 within network 10 may or may not share commonality with other subnetworks. The functional behavior of each node may differ in the number and configuration of connections, weighting of connections, and / or any other aspect. In some edge cases, there may be more than one subnetwork node selection option. In one exemplary edge case, the subnetworks can be defined without the selection option so that the activation of the parent feature results in the activation of the child feature. For example, the parent feature node may connect to one pool and that one pool may connect to one PSCF node.

Nodes in the network are preferably configured to operate, perform, or interact with stochastic interactions that determine node activation, selection, on / off, or other appropriate state. When activated by the parent node, the node preferably triggers the activation of the connected child node according to the node's choice function. Nodes preferably represent binary or multinomial random variables, as in Bayesian networks, but other suitable node models may be used instead. The feature node is preferably a binary establishment variable node that can have multiple parents and multiple children. The parent feature node 110 and the child feature node 150 are considered feature nodes. When multiple parents are involved (ie, multiple nodes are connected via one parent / input connection), the interaction between the parent connections is preferably treated as the superior position of the connection. For example, the child feature node is preferably turned on (ie, selected, activated, etc.) when any parent node is on. Canonical models such as Nosy-OR and Nosy-Max gates can be used to probabilistically model the interaction of multiple parents within a node. The child connection of the feature node preferably encodes the stochastic relationship between the feature and the pool. In a preferred embodiment, if one feature is active, then all pools of that feature are active, but such activation can be modified according to a probability table or any suitable mechanism. As shown in the table below, each link from node to pool node encodes a probability table of type P (pool | feature).

If the feature is on and the pool node is on, then p and q are zero. However, other values of p and q may be used instead. The pool node 120 is preferably treated as a binary node. The pool node 120 preferably has one parent connection representing the probability table described above. The pool node 120 can have a plurality of connections to the child nodes. In one example, a child node connection represents an instant-by-instant connection. For instant connections, it is preferable to execute an OR selection function on the pool element with a related probability. In other words, the instantaneous connection represents a multinomial random variable connection. For example, suppose a particular pool has an Npm pool element. Value 1. .. .. , Consider a binomial random variable M that takes Npm. The departure link from pool node 120 represents the probability distribution P (M | pool). In turn, P (M | pool) defines the probability that a particular pool element will be selected as the starting element of the sequence. Subsequent pool elements are generated in chronological order by following the time selection function (ie, transition function) of that pool element until the end point is reached or the network behavior resolves. Pool elements (also modeled as possible activation sets for PSCF nodes 130) are preferably configured to act as binary random variables, at least one of which is pooled according to distribution P (M | pool). Selected when selected. The pool element represents a combination of functions of child features. For example, suppose pool element 1 is child feature 1 and child feature 2. The constraint node is preferably treated as a binary node whose observed value is instantiated to 1. The probability table used by these constraint nodes implements the kind of constraint that is enforced between the parent nodes that connect to the constraint node. The constraints are often AND or OR constraints, but may be any suitable choice function. These constrained nodes can also be nodes with connections larger than the paired connections.

The parent feature node 110 functions as a higher-level feature node. In the generation operation mode, the parent feature node 110 serves as an input for the subnet 100. In the inference operation mode, the parent feature node 110 is the output of the subnet 100. The parent feature node 110 is configured to execute a choice function upon activation. The selection function is preferably a logical function such as an AND, OR, NOT, or Boolean selection function for node selection. For example, when P1 and P2 are pool nodes of PF1 and PF1 is configured as an AND selection function, activation of PF1 activates the pools of P1 and P2. This choice function may include a randomized choice mechanism for determining the choice between different options, such as when the operator is XOR and only one connected node is selectable. Further, the randomized selection may be biased or weighted according to the node connection weighting of the connection between the parent feature node 110 and the pool node 120. The choice function can, as an alternative, be a probability choice function or any suitable function used in selecting connection options.

The pool node 120 functions as a node for selecting from a set of child features. The child features associated with one pool node 120 are preferably sharing relationships, having correlations, or variations of each other. For example, one pool can be for multiple variations with different pixel pattern positions. In other words, the PSCF node 130 is preferably an invariant representation of a variation of features. In FIG. 3, P1 is an invariant representation of 1 for three different translations of the vertical line, and P2 is an invariant representation of 1 for the three different translations of the horizontal line. As used herein, the term pool may be used to refer to a possible set of PSCF nodes for a particular pool node 120. A possible set of PSCF nodes 130 is preferably any PSCF node 130 having a connection to the pool node 120. The pool may be constrained. For example, a pool element can be a set of {a, b and c, d, e}, where a, b, c, d, e are child features. Like the parent feature node 110, the pool node 120 is configured to execute a choice function upon activation. This choice function can be any suitable function, but is preferably a logical operator as described above for the parent feature node 110. Choice functions can be randomized, biased and / or weighted as well. The selection function of the pool node 120 preferably signals the corresponding PSCF node 130 by selecting, triggering, activating, or otherwise. In addition, the choice function may be restricted or disabled based on the activated constraint node. The activated constraint node may define which node is selected in the pool based on the selection of the PSCF node 130 (connected through the constraint node). Similarly, it can determine the possible set of PSCF nodes 130 for the pool node 120 and / or the weighting or preference of the pool node 120. Pool nodes 120 in the subnetwork can be evaluated sequentially so that constrained nodes can be applied to other pools as needed.

The PSCF node 130 functions as an option for the invariant feature option. One PSCF node 130 maps to one child feature and one PSCF node 130 has only one parent pool node 120. The PSCF node 130 may be further connected or coupled with the constraint node 140. The constraint node 140 preferably defines a relationship between the plurality of PSCF nodes 130. The constraint node 140 preferably connects to another PSCF node 130 with different pools, different times, and / or different subnetworks 100. It is preferred that the PSCF node 130 is not shared between subnetworks. However, the child feature node 150 (which can be the parent node of the lower subnetworks) may share connections to multiple subnetworks.

The constraint node 140 functions to limit the types of patterns allowed in the subnet 100. The constraint node 140 is preferably connected to at least two PSCF nodes 130. Alternatively, more than two PSCF nodes 130 may be connected via one constraint node. The constraint nodes 140 may also be between any suitable type of node. The constraint nodes 140 may be between the pool nodes 120. The constraint node may also be between two types of nodes. For example, the constraint node can connect the PSCF node 130 and the pool node 120. Here, as a preferred implementation example, a constraint node connects PSCF nodes to each other is shown, but the constraint node imposes a constraint between any set of nodes (of any kind) in the network 10. Can be used. The constraint nodes can be between pool nodes, between pool nodes and PSCF nodes, or between any suitable node in the network as shown in FIG. The PSCF nodes 130 are preferably not in the same pool and in some cases are not in the same subnetworks. The constraint node 140 preferably connects PSCF nodes 130 of the same layer to each other, but may instead connect subnetworks 100 of different layers. Further, one constraint node 140 may be connected to any suitable PSCF node 130, or any suitable number of constraint nodes 140 may be connected. Constraint nodes can enforce restrictions, rules, and constraints within the selection of other pools, other subnetworks 100, and / or nodes at different times. The network 10 is preferably evaluated in an ordered manner so that the PSCF nodes 130 connected via the constraint nodes 140 are not evaluated at the same time. When the first PSCF node 130 is activated or selected, the constraint node 140 connected to the first PSCF node 130 is activated. Subsequently, the restriction of the constraint node 140 is activated / enforced on the connected PSCF node. The constraint node 140, like any other node, may have a choice function that determines how to activate the PSCF node. The constraint node 140 preferably affects how the pool node 120 can select the PSCF node. In one example, the choice function of constraint node 140 is an AND logical operator that performs the selection when one of the connected PSCF nodes is active. In another variant, the choice function of constraint node 140 can be an OR logical operator that modifies possible PSCF nodes in the pool. You can use any suitable choice function. Some constraint nodes 140 may have a basic or simple constraint that activation of one node corresponds to the selection of a second node. Since the selection logic is direct correspondence between nodes, these can be represented as direct connections without nodes. Preferred variants of the constraint node 140 may include a lateral constraint node 142, an external constraint node 144, and a time constraint node 146.

The lateral constraint node 142 functions to limit the types of patterns in the subnetworks based on the interaction between the pool nodes 120 in the subnetworks 100. The lateral constraint node 142 is preferably an enforcement rule or node connection between the PSCF node 130 of the first pool node 120 and at least one second PSCF node 130 within the second pool node 120. These first and second pool nodes share a common parent node 110. Lateral constraints are used so that the various configurations generated correspond to the horizontal translations of the corners. In this case, the parent feature can be considered to represent various variants of the corner. That is, the parent feature has a corner representation that is invariant to the corner deformation.

As shown in FIG. 4, exemplary implementations of the subnetwork are two pool nodes 120 (P1 and P2) and three six PSCF nodes 130 (three for each pool node 120). X1, X2, X3, X4, X5, X6), three lateral constraint nodes 142 (C1, C2, and C3) connected to a pair of PSCF nodes 130, and individually connected to one PSCF node 130. It has six child feature nodes 150 (CF1, CF2, CF3, CF4, CF5, CF6). The parent node's choice function is an AND operator such that both P1 and P2 are selected at the same time. When the parent feature is selected, the pools P1 and P2 connected to the parent feature node are automatically selected. Both P1 and P2 are configured as XOR logical operators for a particular time instant. Only one offspring (ie, connected PSCF node 130) is selected as the XOR operator at a time. In a mode in which each pool node P1 and P2 randomly selects one of its PSCF nodes, the lateral constraint nodes 142 and their connections encode the constraints imposed between child feature selections of different pools of the same parent. .. In this implementation example, C1 and C2 are AND operators. Therefore, when X1 is activated, X4 is forcibly activated by the connection of C1. Similarly, when X2 is activated, the connection of C2 forces the activation of X5. C3 can be implemented as an XOR. When X3 is activated, only one of X3, X4, and X5 can be selected, so the connection of C3 between X3, X5, and X6 prevents P2 from selecting X5 and X6. These serve only as exemplary connection and selection functions. Each node can have any suitable function and connection architecture. In this example, if there is no lateral constraint node 142, each pool can independently select child features, and the parent features are one feature from P1 and another feature independently from P2. Corresponds to nine different patterns of invariant representations for all of the combinations that can be generated by selecting. In some subnetworks, the constraint node 140 may not be used.

The external constraint node 144 functions to enforce an immutable pattern across different subnetworks 100. Similar to how lateral constraint nodes 142 ensure that representations in different pools are consistent with each other by imposing a constraint on which PSCF node 130 of one pool node 120 can be combined with PSCF nodes of other pools. In addition, the external constraint node 144 can maintain the compatibility of the entire hierarchy. External constraint nodes 144 preferably create connections, rules, or other constraint mechanisms that can generate selective interactions between two different subnetworks 100. The external constraint node 144 preferably connects at least two PSCF nodes 130. Like other constraint nodes, the choice function is executed when activated. As shown in FIG. 5, the layered network 10 can be configured with at least one upper layer subnetwork 100 that provides inputs to at least two lower layer subnetworks 100. When generating a sample from the network, the generated features of the top layer subnetwork 100 are used as inputs for simultaneously generating a sample from the lower layer subnetwork 100. Without external constraints, the samples produced by pool 1 and pool 2 will not be coordinated with the samples generated by pool 3 and pool 4. However, the external constraint nodes shown in bold in FIG. 5 provide a mechanism for performing coordination between pools of subnetworks 100 in the lower layers. As shown in FIG. 5, PSCF nodes 130 can have more than one type of constraint node forced on them. Lateral constraint nodes 142 impose coordination between PSCF nodes 130 in different pools of the same network, and external constraint nodes 144 impose coordination between PSCF nodes 130 in different subnetworks 100. The constraint node 140 is preferably configured so that there is no contention (eg, if one constraint activates the node and the other specifies that it should not be activated). Conflicts and competition between constraint nodes 140 can be resolved using heuristics regarding the ranking of constraint nodes 140, or the order in which constraint nodes 140 are implemented, or other appropriate rules.

The time-constrained node 146 functions to implement the relationship between the network 10 and the subnet 100 operating at other time zones. At the base level, each element of the pool (eg, PSCF node 130 with a shared parent pool node 120) can have a relationship that specifies the order of time they occur. The time constraint node 146 is preferably a simple direct connection constraint, where activation / selection of one node results in selection of the designated node in the second instance. The time constraint 146 does not have to have strict ordering. The time constraint node 146 can specify a set of possible pool elements that can occur in the second moment, given the pool elements that occurred in the first moment or a set of pool elements. In another description, the constraint node 140 may perform a specification-like function in a Markov chain. As shown in the exemplary time constraint of FIG. 6, when the PSCF node "a" is activated at time t, it proceeds to the PSCF node "b" at time t + 1, and the PSCF node "b" at time t becomes the PSCF node "b" at t + i. Proceed to "c". For the first pool, this example uses the representation of two time slices to represent the sequence from "a" to "b", "c". For the second pool, the time constraint node defines a sequence from "d" to "e", "f". The choice function of the time-constrained node 146 can be arbitrarily complicated like any other node. For example, when node a is active at time t, node b or c can be active at time t + i. Higher-order temporal relationships can be further used. While the time-constrained node 146 preferably defines a relationship that advances to a later moment in time, the time-constrained node may define a relationship between a plurality of instances. For example, a time-constrained node can define a constraint from one PSCF node to the network at three different moments. In addition, the time constraint node 146 in some variants can even define a retroactive constraint to the previous moment. For example, inferring the detection of one pattern at this time can enhance the queue for detecting patterns at a previous time point. Here, one instance of network 10 is preferably the operation or use of this network 10 in different time instances. The time-constrained node 146 can be particularly useful for generation or inference applications to time-based media such as video, audio, or computer graphics. Although time-constrained nodes are specific to time-based applications here, it is possible to define similar constrained nodes 140 between different instances of network 10 along other dimensions depending on the area of challenge or use case. it can.

As shown in FIG. 7, the network 10 can be configured for inference or recognition use case scenarios. When the network 10 is configured for inference, the propagation of node selection flows upwards in the hierarchy from the data feature (ie, the lowest child feature), causing the network 10 to operate in an essentially similar but opposite manner. .. Instead of moving from higher-level features to more detailed features, we use detailed features to infer the general functionality of the upper layers. In application, inference can be used to take image features (image properties or sub-image components) and extract information based on those features. The inference configuration network 10 preferably uses the posterior distribution of the nodes (ie, the probabilities of the parameters given the evidence) and the evidence supplied to the child nodes to layer the activation, selection, and on / off states. Propagate to the upper level of. This can be characterized as a variant of the probability propagation algorithm that can be used to derive an approximation of the posterior distribution at the node of interest using local message propagation. Since it is preferable that all nodes in the network are treated as binary values, after the node, it is a mechanism for the node to specify the probability that the node will be turned ON or OFF when evidence (child node) is given. The nodes are preferably configured to pass messages through the illustrated connection channels between the nodes. These connections can be bidirectional conduits for messages. The messages flowing upstream are the likelihood message and the probability message downstream. As shown in FIG. 7, the subnet network propagates the message based on the input image. The message in this example represents the likelihood of evidence, assuming that the node corresponding to the source of the message is on. Therefore, node CF2 has a higher likelihood than node CF1 because the representation of CF2 is more consistent with the input evidence. The likelihood of a pool is the maximum value over the likelihood of each pool element. When a network has a set of inputs corresponding to subsequent time instances, the network can propagate messages in time and make temporal inferences. In that case, the evaluation of different nodes is an expression of the probability given a set of evidence.

As shown in FIG. 8, the network 10 can further include at least two subnetworks 100 having overlapping, overlaying, or shared child feature nodes, which interact with multiple parents. Works to introduce. The shared child features of the two subnetworks preferably include two PSCF nodes of two different subnetworks with child connections to the same child feature node. During inference, such network architectures can produce descriptive effects. For example, the feature node "b" shared by two subnetworks can be used in the Nosy-OR mechanism model so that the two subnetworks can compete for the evidence presented to that node. Conflicts are enabled through a belief propagation mechanism or a similar message propagation mechanism. Subnetworks are preferably configured individually, with each subnet having a separate representation of pool elements, constraint connections, and PSCF nodes. Child features that participate in two different parent features as part of two different subnetworks can have different activation values during pattern generation and inference. As shown in FIG. 9, a network that can be characterized as a recursive cortical network when used recursively can form any large and complex network 10.

2. 2. Method of Creating a Neural Network [0036] As shown in FIG. 10, a method of creating a neural network S10 of a preferred embodiment is a step S100 of recursively constructing a plurality of sub-networks in a network hierarchy, and is a first layer sub. A step including step S110 to join a child feature node of the network with a parent feature node of the second tier subnetwork; and a step of setting a selection function for the parent feature node connected to at least two pool nodes in the subnetwork. With S120; with step S130 setting the selection function of pool nodes connected to at least two parent-specific child feature nodes (PSCF nodes for short) in the subnetwork; at least a pair of PSCF nodes via constraint nodes. It includes step S140 to link; and step S150 to propagate the network selection down the hierarchy of the network layer in a manner consistent with the node connection of the subnetwork and the node selection function of the subnetwork. This method preferably functions to construct, create, manufacture, or transform a neural network into a usable system. The network created by method S10 further uses coordinated connections between lower and upper levels, selectivity by constraint nodes, and shared learning through a sparsely distributed representation of features. , Acts to enhance immutability. The completed network is preferably substantially similar to the system described above, but any suitable modifications or alternatives can be incorporated into this method S10. When used for generation, output feature nodes are preferably connected to assemble multiple features into a generation pattern. In one implementation example, the output of multiple child feature nodes is assembled into a generated image. In another implementation, the output of multiple child feature nodes is assembled into a voice signal. When used for inference, the child feature node receives data input. In this implementation example, the calculated image features derived from the image processing technique are supplied to the child feature nodes. Similarly, the input data can be voice, data signals, or any suitable data feature.

Block S100, which includes the step of recursively constructing a plurality of subnetworks in the network hierarchy, functions to reuse the subnetprint patterns in the layered network. Subnetworks preferably interface with each other via feature nodes, i.e., parent feature nodes and child feature nodes. The parent feature node is preferably at the top of the network hierarchy, and the connection branches from the parent feature node to the child feature node in the end. Subnetworks preferably have several possible child feature nodes. This is a set of nodes that are the leaves of the network, or, in other words, lower level / layer nodes. Here, a node can be described as any suitable description of a node in an artificial neuron, artificial neurodes, processing element, processing unit, or artificial neural network. The node is preferably an operating element that includes a parent connection for the node to receive a signal for activation and a child connection for the node to send a signal to the node connected for activation. Activation can be further described as selection, node configuration state (eg, on or off state), or any suitable output. The activation signal is preferably binary, but may have any suitable number of states. Subnetworks preferably include parent nodes, pool nodes, PSCF nodes, optionally constraint nodes, and child features, but such node classifications and types may be adjusted appropriately. For example, PSCF nodes may be functionally coupled with child feature nodes when configured into physical logical blocks within the processor. Subnetworks are preferably configured individually, with each subnet having a separate representation of pool elements, constraint connections, and PSCF nodes. In other words, each subnet network is often not the same as each subnet network, but instead includes customized connections, number of nodes, constraints, and other individually configured configurations. However, it is preferable that the subnetwork architectural pattern is consistent within the subnetwork. Node and connection settings are preferably automated via training data, live data, or historical data. Node configuration can also be semi-automated with adjustments and customizations via user input. The network settings may be fixed or updated continuously or periodically. Any suitable number of layers of the subnetwork can be used. In addition, recursively designing multiple subnetworks means designing at least a second network for the second time instance, and designing subnetworks with overlapping, overlaying, or shared child feature nodes. Can include.

[0038] The block S110 including the step of connecting the child feature node of the first layer subnetwork with the child feature node of the parent feature node of the second layer subnetwork connects the subnetwork through the parent feature node and the child feature node. It works to do. Subnetworks are preferably organized in multiple layers. Subnetworks of the same tier preferably have the same parent subnetwork, or at least a parent node connected to a subnetwork in the same tier as the parent subnetwork. The subnet hierarchy preferably starts at the highest level with open parent feature node ports and extends downward to the lowest level with open child feature node ports. Subnetworks can be separated into separate layers in which one "child" subnetline is descending from one "parent" subnetworks. However, in one variant, the subnetworks can be configured to interact as subnetworks of any layer. It also has an intermediate special case subnetwork with a simple network where the choice function is an identification function with 1 parent node, 1 pool node, 1 PSCF node, and 1 child node. Can also be achieved by. When the parent node becomes active, the child node becomes active. Any suitable number of layers and any number of subnetworks can be configured within a single layer. In one variant, as shown in FIGS. 8 and 9, at least one child feature node of the first subnetworks can be shared with the second subnetworks in the same layer. In addition, this method can include setting up a posterior distribution model of the nodes within the node, which serves to allow belief propagation for inference or detection use cases. The posterior distribution preferably provides the probability that a node will be activated (ie, turned on) given evidence (eg, a child node).

Block S120, which includes a step of setting a selection function for a parent feature node connected to a pool node in the subnetwork, functions to define and configure activation of the pool of parent features. Preferably, the parent feature node selection function is an AND function that selects all connected pool nodes. Alternatively, the selection function can be an XOR function (selecting only one pool), an OR function (randomly selecting at least one pool node), or any suitable logical operator function. In addition or alternatives, stochastic modeling can be incorporated into Boolean logic functions, stochastic choice functions, or other suitable choice functions. Different pool nodes can receive enhanced / priority weighting or deprecated / reduced weighting. Pool nodes may be selected using a random selection mechanism in conjunction with stochastic modeling. The choice function can be an operation instruction encoded in a digital medium. Alternatively, the operating instructions may be encoded into a physical processor gate architecture.

Block S130 includes, within the subnetwork configuration, a selection function for pool nodes connected to at least two PSCF nodes, and features related, related, or otherwise related to child features of the parent node. Serves the function of defining and configuring node activation. The choice function is preferably configured to be triggered by activation by the parent feature node. Preferably, the pool node selection function is an XOR function that selects one of the child PSCF nodes. The choice function can be any suitable alternative function as described above. One or more PSCF nodes can be selected based on the behavior of the selection function. A pool of PSCF nodes is used to set the invariant pattern of a group of features. Depending on the layer of the subnet, the immutable pattern can have any suitable level of abstraction. For example, a pool of PSCF nodes in the lower layer may accommodate different deformations of a rectangular pixel pattern along a dimension. An exemplary upper layer may include a pool of PSCF nodes corresponding to the legs of different types of animals. This PSCF node can accommodate any suitable invariant pattern collection.

Block S140 includes the step of linking at least a pair of nodes through constraint nodes and functions to define the interaction between isolated parts of the network. The constraint nodes are preferably between at least two PSCF nodes, but may instead be between any set of nodes. Linking PSCF nodes through constraint nodes preferably allows the selection of one node to affect and change the behavior of the selection function of the other pool. As mentioned above, the three preferred types of constraint nodes include lateral constraints, external constraints, and temporal constraints. These constraints are the preferred form of constraints on data in spatial and / or temporal form. If the alternative network or subnetworks occupy another data dimension, other forms of constraints can be used additionally or alternatives. Constraint nodes can further include a choice function, as described for other types of nodes. The choice function can be set in the same way. You can configure any suitable number of input and output connections to the constraint node. In a basic implementation, the constraint node is an AND logical function that performs the selection of the connected PSCF node when the connected PSCF node selects or activates the constraint node. The constraint node can include a predefined input connection that defines which PSCF node is used to activate the constraint node. Alternatively, the constraint node may use all connections of the PSCF node in mixed mode, in which case the first signal indicating activation or selection triggers the constraint node and the remaining connected PSCF nodes. Enforce selection constraints on (the pool node has not yet selected the PSCF node).

In one variant, block S140 includes the step of linking the first node of the first pool to the second node of the second pool, where the first pool and the second pool are Share the same parent feature node in the same subnetwork. Such constraints between sister pool elements are preferably defined as lateral constraints. The lateral constraint nodes are preferably between at least two PSCF nodes, but may instead be between any set of nodes. In another variant, block S140 comprises linking the first node of the first subnet network to the second node of the second subnet network, where the first subnet network and the second subnet are linked. The network is a different subnetwork. Such constraint nodes between different subnetworks are defined as external constraint nodes. The first and second subnetworks are preferably in the same tier within the network, but may be in different tiers. Similarly, the external constraint node is preferably between at least two PSCF nodes, but may instead be between any set of nodes.

In another variant, block S140 includes the step of linking the first node of the first network to the second node of the second network, where the first network is the first instance. (Eg time t) is specified and the second network is specified for the second instance (eg time t + 1). Such a constraint node between different times is a time constraint node. The first and second instances are equivalent networks but evaluations in two different time periods. The time-constrained nodes are preferably between at least two PSCF nodes, but may instead be between any set of nodes. The first and second instances may also be evaluations of two different networks assigned for continuous network evaluation. In yet another variant, the first and second instances may be defined through child features (where one feature set is for one period and the second set of features is for different periods. is there). In this variant, the time constraint can also be considered a lateral or external constraint. In addition, other forms or types of constraint nodes can be used, depending on the specific use case in which the underlying network architecture is used.

Block S150 comprises propagating node selection through the network hierarchy in a manner consistent with the node connection of the subnetwork and the selection function of the nodes of the subnetwork, and functions to activate the network by input. The network can be used for pattern generation and / or pattern inference. Pattern generation preferably uses upper layer or vague pattern inputs to transform, project, or synthesize new or potential objects or interpretations. Pattern inference preferably transforms data that represents physical evidence of an object, event, or metaconcept into a pattern interpretation that is revealed through propagation over the network. The generation mode and estimation mode may additionally, at least partially, use other modes for improving, enhancing, or facilitating the generation or inference of pattern output. Similarly, one mode of operation can be used to enhance the stochastic model of nodes in the network. Preferably, the training data is propagated iteratively or continuously through the system using a training engine that updates the probabilistic model according to any suitable heuristic, algorithm, or approach. Propagation of node selection is preferably performed as described in the following method, but any suitable technique may be used instead.

As shown in FIG. 11, an exemplary implementation of method S10 is a hierarchy of three subnetworks and nodes set using their respective choice functions. The parent node is set with an AND-related function represented by a solid connection arrow so that each connection is selected at the same time. When you select a parent feature, the pool node connected to the parent feature node is automatically selected. The pool node is set by the XOR relational function represented by the dashed connecting arrow. At a particular point in time, only one descendant can be selected. In addition, this function uses a random selection process to select one of the PSCF nodes. The three lateral constraint nodes use the AND function so that when one of the PSCF nodes is selected, the corresponding node in the other pool is selected.

3. 3. Method of Generating Patterns from a Network As shown in FIG. 12, method S20 of generating patterns from a network in a preferred embodiment is a recursive subnetwork having a parent feature input node and at least two child feature output nodes. Step S210 to provide the network; and step S220 to propagate the node selection through the network hierarchy in a consistent manner with the node connections of the sub-networks of the network, and in the parent feature node, match the function of the parent feature node. With step S230 selecting the pool node; with step S240 selecting at least the first PSCF node corresponding to the child characteristics of the subnetwork in the pool node; at least the second in response to the selection of the first PSCF node. Includes step S250 to enforce selection constraints on the PSCF node of the network; and step S260 to compile the final child features of the network into generated output. The method S20 functions to synthesize, simulate, or generate patterns based on network patterns. The network is preferably a neural network or a Bayesian network as described above, and can operate on an arithmetic device. Method S20 preferably transforms the high-level inputs of the top-level parent feature node into detailed data features that can be assembled or combined from the outputs of the child feature nodes. This method is preferably performed on a single instance of the network, but method S20 may be further extended to operate periodically or continuously for multiple instances. Similarly, this method can include any suitable adjustments so that one network can operate in concert with additional networks. Pattern generation can be applied to various media and disciplines such as computer graphics, speech synthesis, physical modeling, data simulation, and natural language processing / translation. In one implementation example, method S20 can be used to generate an image based on contextual information. In other embodiments, method S20 can be used to produce a synthesis. Pattern generation can be modified for use in predictive-based applications. Predictions can be seen as a special case where the generated content is projected into future time. Such implementation examples may include financial trend forecasting or data analysis.

Block S210 includes a step of providing a parent feature input node and at least two child feature output nodes in the network of the recursive subnetworks, and functions to implement a recursive cortical network with enforced constraints. To do. This network is preferably a network as described above or as created by method S10. The base network of any complexity preferably has at least two layers with one top layer subnetting and two lower layer subnetworks. The network of recursive subnetworks is preferably more complex with multiple layers within the hierarchy. Each subnetwork can have any suitable number of child feature nodes, which can be used by any number of progeny / child subnetworks as lower layer inputs. For example, a first-tier subnetworks can have five pools, each with 2, 3, 4, 5, and 6 connected child nodes. Having a total of 20 child nodes, the second layer can have 20 different subnetworks that use those child feature nodes as inputs to their respective parent feature nodes.

Block S220 includes a step of propagating node selection through the network hierarchy in a manner consistent with the node connections of the network's subnetworks, to set selection, activation, on or off, or the state of the nodes in the network. Works for. The step of propagating the node selection can include sending an electrical signal that acts as a trigger or activator to derive the proper activation of the target node. The step of propagating a node selection can also include the step of sending a message or communication to another node. Protocols may be in place to coordinate communications / messaging. Propagation of node selection preferably involves a systematic or systematic approach to node activation. First, the pattern's parent feature input is received. The parent feature is preferably another input on which to base the higher feature, classification, or pattern generation. The input is preferably supplied to the top subnetwork of the network. Propagation over the network then proceeds: the top layer subnetwork is processed; then the next layer of the subnetwork is processed; and processing continues, progressively at each layer of the network. Processed (ie, sequentially or continuously). In another variant, at least a partial child feature input is received at the lowest child feature node. This variant includes a step in which the network performs generation within some of the supplied child feature node inputs. This variant includes the step of receiving a seed child feature input (eg, at least a partial selection of child feature nodes) that serves to provide a context and framework for generation. For example, once half of the image is fed, propagation within the network is used to generate / create / imagine child features for the other half of the image. In special cases where there is a single layer (eg, a set of sister subnetworks), the propagation of node selection is preferably done in parallel and / or in sequence across the subnetworks, depending on the configuration. In addition, there can be an order of processing of subnetworks within a single layer. In some examples, the external constraint defines the relationship between the two subnetworks, one subnetwork is processed first, then another subnetwork is processed with the external constraint in mind. .. This order may be predefined or configured. Alternatively, the process may be a race condition between the different subnetworks and the first subnet network to complete the process of determining the enforcement of the constraint. Alternatively, they may be processed or operated simultaneously in any suitable manner. Similarly, the processing of nodes in the subnetwork may be in order. It is preferred that the pools in the subnetworks be ordered in the same way. In some examples, the lateral constraint defines the relationship between the PSCF nodes of the two pools, one pool being processed first, then the other pool being processed with the lateral constraint in mind. This order may be predefined or configured. Alternatively, the process may be a race condition between a different pool and the first pool to complete the process and determine the enforcement of constraints on other pools. Alternatively, they may be processed or managed simultaneously in any suitable manner. Within each subnet, blocks S230, S240, and S250 are preferably implemented. Node selection preferably begins with the parent feature node, then the pool node is activated, and the PSCF node is selected. The choice of PSCF node may be determined at least partially affected by the enforced selection constraint of the constraint node.

Block S230 includes the step of selecting at least two pool nodes that are consistent with the function of the parent feature node and functions to properly activate the pools in the subnetworks. As mentioned above, the pool is preferably a group of PSCF nodes corresponding to invariant features. The selection is preferably made within the parent feature node configured by the selection function. The choice function is preferably an AND function such that each connected pool node is activated, but any suitable choice function can be used instead.

[0050] Block S240 includes selecting at least a first PSCF node that corresponds to a child feature of the subnet, and functions to select a PSCF node within the set of pool elements of the pool node. This selection is preferably made for each of the selected pool nodes from block S230. The order in which pool nodes in a subnetwork are evaluated can be random and non-simultaneous. Alternatively, multiple pools may be evaluated at the same time. The selection of PSCF nodes is preferably performed according to the selection function of the selected pool node. In one embodiment, the selection function is an XOR function and only one PSCF node is selected. Alternatively, any suitable choice function may be used. One PSCF node is preferably connected or associated with at least one child feature node in a direct relationship, and when this PSCF node is selected, the connected child feature node is selected. In some variants, the PSCF node may be associated with multiple child feature nodes. Each child feature node is preferably selected when the corresponding PSCF node is selected. In yet another variant, the child feature node may also be associated with other PSCF nodes in the network or subnetworks. The child feature node is preferably selected / activated based on the superior position of the connection to the child feature node.

Block S250 comprises the step of implementing a selection constraint on at least the second node and functions to make it possible to define an invariant relationship between the pool and the subnetworks. Constraints are preferably created to define the logic between feature combinations and patterns. In a typical example, subnetworks stitch together multiple image components to form an image of a car, limiting car wheels to the other pool when one pool selects the car body. Can be enforced, which maintains the consistency of the wheels and the car body. Selection constraints may be defined via a connection between at least two PSCF nodes via constraint nodes. The constraint nodes include any suitable number of connected PSCF nodes and can execute any suitable choice function. In some variants, selection constraints may be defined via connections between two pool nodes or any suitable type of node. Similarly, the constraint node may be between any two or more types of nodes, such as between a PSCF node and a pool node. Execution of the constrained node preferably has some form of direction at run time, and the choice of the first node has the effect of the choice on the second node. The directionality can also go in any direction between the two types of nodes. The PSCF node can result in the constraint node affecting the pool node, and the pool node can result in the constraint node affecting the PSCF node. One preferred selection constraint would be to force the selection of the connected PSCF node if one of the PSCF nodes connected to the constraint node is activated. In other words, the selection constraint function of the constraint node is an AND operation. The selection constraint is preferably implemented in response to the selection of at least a first PSCF node having a connected constraint node. As mentioned above, the nodes are preferably evaluated or propagated in some order. Selection constraints are preferably not enforced on PSCF nodes that have already been selected, but instead apply to selection by pool nodes. In some scenarios, the pool node can reduce a set of possible PSCF nodes to one node after the selection constraint is enforced and sent to the pool element through the constraint node. In other scenarios, pool nodes can reduce the number of possible PSCF nodes or change the stochastic weighting for selection. The constraint node is shown as a connection between two PSCF nodes, but as an alternative, constraints can be implemented via a message passing mechanism between pool elements and / or subnetworks. These messages preferably modify the behavior of the choice function to effectively execute the constraint node as described in this document. Constraint nodes can be horizontal constraints, external constraints, time constraints, and / or any suitable type of constraint. Lateral constraints are preferably implemented between two different pools. External constraints are preferably implemented between two different subnetworks. Lateral and external constraints are preferably used for spatial constraints, but may be used to define any suitable invariant pattern. Time constraints perform network evaluations on different time instances. Time constraints can define invariant patterns over different time frames. The temporal selection constraint determines which features may or may not occur within a set of features.

Block S260 includes the steps of compiling the final child features of the network into the generated output, and the plurality of features are generated into one artifact, representation, or analysis, imitation, or any suitable output. Works like assembling. The final child feature is preferably the child feature node at the bottom of the hierarchical network. These child feature nodes preferably represent binomial variables that represent the presence of a particular data feature. You can maintain a database or mapping that maps child feature nodes to specific data features. As shown in the example of FIG. 4, the child feature nodes CF1, CF2, and CF3 individually represent vertical bars in different horizontal positions. The step of compiling the final child feature preferably involves mapping a plurality of selected child feature nodes to the data feature, which is then compiled into the generated output. Activated child feature nodes are preferably components that, when combined, form a reproduction of a medium. Preferably, the output is similar to that of the data medium used to train or create the network. For example, if the network is trained or created for image generation, the output is preferably a substantially perfectly mimicking image. If the network is trained with audio features, the final child features can be combined to output an audio file or signal. When multiple network evaluations are used for time signals, the final child features of multiple networks can eventually be compiled into the generated output. In an exemplary embodiment, the final child feature is associated with a spatial element. In other words, each child feature corresponds to a particular feature in a particular aspect. In the case of images, this spatial component is preferably a two-dimensional block position. Spatial components may further be three-dimensional or arbitrary appropriate dimensional data. The dimensions may correspond to physical or artificial dimensions. For example, when used with data analysis, each child feature may correspond to different dimensions of data analysis, which can preferably be compiled to form the appropriate output. Since the child feature node is preferably a node in the binary activated state, there is preferably a mapping between the child feature and the associated feature pattern. It is preferred that the child features uniquely correspond to a particular aspect of the medium. For example, for an image, the image can be divided into a grid of blocks, and each child feature node in the lower layer can be associated with a pixel pattern (eg, a 3x3 pixel pattern). By this method, the activated node is used to select various pixel patterns, and those pixel patterns are superposed at appropriate positions in the final image.

4. Method of Inferring a Pattern from an Input Using a Network [0053] As shown in FIG. 13, a method S30 of inferring a pattern from a network of a preferred embodiment has a parent feature input node and at least two child feature output nodes. With step S310 to provide a network of recursive subnetworks; with step S320 to configure nodes of the subnetwork using a posterior distribution model; node selection through the network hierarchy in a way that matches the node connections of the subnetwork of the network. With step S330 propagating; with step S340 executing selection constraints on at least the second PSCF node in response to the selection of at least the first PSCF node; with step S350 outputting the parent feature node of the network to the estimated output. including. Method S30 functions to infer or detect patterns in the input data. Input data is preferably supplied during inference. Data is transformed, processed, or modified into data features. Then use the data function to selectively activate / select different child nodes. Probabilistic propagation or similar message passing approaches are then implemented on the network, with node selection from the lower subnetwork (eg, the layer of raw data features) to the upper subnet (eg, the layer of pattern features). It is done to work. Method S30 can be used for inference patterns of a wide variety of data types such as image, video, audio, speech, medical sensor data, natural language data, financial data, application data, traffic data, environmental data and the like. In one implementation, this method can be used for image detection to detect the presence of an object in an image or video. In addition, this method can be used to detect multiple objects within the same image.

Block S310 comprises providing a parent feature input node and at least two child feature output nodes to the network of recursive subnetworks, and functions to implement the recursive cortical network with enforced constraints. .. The network is preferably a network as described above or as created by method S10. Since the architecture of the network is similar to that used in method S20, methods S20 and S30 can use the same network configuration for either generation or inference use cases.

Block S320 functions to provide a probabilistic model from which the inference can propagate up the hierarchy, including the steps of constructing the nodes of the subnetwork in the posterior distribution model. Inference is the process of finding a posterior distribution on every node in the network, given some evidence on a subset of the nodes. The posterior distribution model is preferably a parameter that specifies the probability that the node will be activated given the evidence. The evidence is preferably a set of nodes that are children of a particular node. The posterior distribution model can include a posterior distribution of a set of evidence possibilities, where the set of evidence possibilities includes a permutations of evidence nodes that are ON and OFF. The posterior distribution can be expressed as a probability, a ratio, a logarithm of the ratio, a weighted choice function (if the node is active depending on the evidence), or any suitable representation. Alternatively or additionally, other inference mechanisms may be incorporated into method S30.

Block S330 includes a step of propagating node selection through the network hierarchy so as to be consistent with the node connections of the network's subnetworks, and functions to propagate probabilistic reasoning above and below the network hierarchy. Block S330 preferably uses belief propagation, but other probabilistic inference approaches may be implemented as alternatives. Belief propagation is preferably used to propagate selections above the network hierarchy. Propagation of node selection is also consistent with the posterior prediction of child nodes. Belief propagation involves exchanging messages between nodes and performing calculations within the nodes under different assumptions. Links between nodes can be constructed as bidirectional communication channels for messages. In one implementation, the message flowing upstream represents the possibility and the message flowing downstream represents the probability. In addition, generation or down-propagation can be used to enhance and provide feedback on up-propagation. For example, the generative propagation described above can be used to imagine what could be in the image and compare it to what is actually in the image. In the example shown in FIG. 7, the message propagating the network link can be used to infer the pattern of the image. The message in this example represents the likelihood that evidence exists, assuming that the node corresponding to the source of the message is on. For example, node CF2 has a higher likelihood than node CF1 because the representation of node CF2 is more consistent with the input evidence. The likelihood of a pool (represented by connections originating from a pool node) is the maximum likelihood of a pool element. When propagating a probability to a network using a set of inputs corresponding to subsequent time instances, the network can propagate the message in time and make temporal inferences. In such a scenario, the values calculated at different nodes represent the probabilities given a set of evidence.

[0057] Preferably, propagation is initiated when a data feature input is received at the last child feature node of the network. The last child feature node is the lowest child feature node in the hierarchy. Preferably, the data is processed, transformed, or divided into a set of features. The data feature is then used to select or activate the last child feature node. In a simple scenario, the presence of the feature is used to activate or deactivate the child feature node. Alternatively, the likelihood parameter of the feature node can be input. Likelihood can be a convolutional similarity measure, or any suitable measure in which features may be apparent in the data. Belief propagation continues to propagate this input up the network hierarchy. Within the subnetwork, the step of propagating node activation is the step of sending the likelihood score to the PSCF node to which the child feature node is connected; in the pool node of the subnetwork, of the PSCF node connected to the posterior distribution component. The step of generating the likelihood score from the likelihood score; and the step of generating the likelihood score from the posterior distribution component and the likelihood score of the pool node connected to the parent feature node in the parent feature node of the subnetwork. including. Belief propagation then preferably continues to the higher subnetworks until network propagation ends or some threshold is met.

[0058] Block S340 includes at least a step of implementing a selection constraint on the second node and functions so that the invariant relationship between the pool and the subnetworks can be defined and used during inference. It is preferred that the constraint nodes and connections be implemented in a manner substantially similar to that of block S250. When a node is activated, the constraints applied to other nodes connected through the constraint node are enforced. The external constraint node is preferably between at least two PSCF nodes, but may instead be between any set of nodes. In one variant, the constraint can, as an alternative, increase or change the probability measure of the connected PSCF node and / or the PSCF node of the same pool.

[0059] Block S350 includes a step of outputting the characteristics of the parent feature node of the network to the inference output, and functions to process or assimilate the activated node of the network into the inference result. Preferably, the parent feature node is used as an indicator of the pattern. When building a network, preferably different layers detect patterns on different granularity scales. At lower levels, it involves detecting specific pixel patterns such as corners, lines, and points. At a high level, it can be a pattern detection, such as a person being detected in an image or a message expressing happiness. Also, each subnetlight is preferably customized for specific pattern recognition. In the above example, the subnetworks can be for invariant corner detection. When the parent node of this particular subnetworks is activated, it can be inferred that a corner exists. Mappings may exist so that the activation of the parent node of the subnetwork is paired with a separate pattern label. Inferences may come from the top layer, but instead may be obtained through multiple layers of the network. For example, if this method outputs the inference that "a male human is laughing", then the inference that there is a human, this human is a male, and the facial expression is a smile is in multiple layers and / Or obtained through a subnetwork. In addition, the scope of inference can be adjusted by selecting which layer and / or subnetwork to use when outputting inference. For example, when generating inference from an image, the inference from the upper layer can detect that the image is from a coffee shop scene. The lower layer can be used to detect the presence of three tables, men, women, and various other coffee shop fixtures in the image. APIs or interfaces may be present to extract the appropriate level of inference information from the network. This interface preferably responds to the request for analysis by appropriately selecting and formatting the response with appropriate inference information. For example, in the interface, the request could be "How many people are in this image", "What is the background of this scene", "Where is the swan in this image", or "Next to the object in this scene". It is possible to identify a specific pattern of interest, such as "what is the state?" The more open-ended requests, the more patterns can be detected. The inference information specified in the set of activated parent nodes in the subnetwork can be used in any suitable way instead.

5. System and Method Use Examples [0060] As mentioned above, recursive cortical networks can be used in a wide variety of scenarios. The network architecture is useful for both generation and inference. Networks are also preferably agnostic to the format of the data used as input for generation or inference. Preferred data media include 2D or 3D image data, image sequences, video, audio, natural language text, analytical data, widely distributed sensor data, or other suitable format data. In one preferred application, the system and method are applied to the image. The network can be used in inference mode for object detection, event analysis, face recognition, mood detection, object tracking, and other suitable uses. In generation mode, the network can generate simulated images. As another exemplary application, the system and method can be applied to natural language processing. Inference mode can interpret sentence context and intent, translate languages, and detect other language patterns. For example, the meaning of a question can be interpreted and then used in network generation mode to generate a response or response. In some implementations, this response is a natural language response, but it may also trigger an action or event that goes beyond mere verbal communication. Other exemplary applications include medical scanning and in-image anomaly detection, financial data analysis and prediction, advertising targeting, traffic prediction, environmental simulation, and other suitable areas of simulation, detection, or prediction. .. Networks are preferably created for specific applications and use cases (such as networks dedicated to image analysis), but networks allow for more general and higher levels of patterns and functionality through the extension of the node's network. Can be used in further combinations. While the network is in use, the network configuration is preferably updated and expanded to take into account new data, which further extends the processing power over the network.

[0061] The systems and methods of preferred embodiments, and variations thereof, can be implemented and / or implemented, at least in part, as devices configured to accept computer-readable media for storing computer-readable instructions. .. Instructions are preferably executed by computer-executable components integrated with a recursive cortical network. The computer readable medium can be stored in any suitable computer readable medium such as RAM, ROM, flash memory, EEPROM, optical device (CD or DVD), hard drive, floppy drive, or any suitable device. The computer-executable component is preferably a general or application-specific processor, but any suitable dedicated hardware or hardware / firmware combination device may perform alternative or additional instructions.

As those skilled in the art will recognize from the above detailed description and drawings and claims, the present invention will not deviate from the scope of the invention as defined in the appended claims. Modifications and changes can be made to the preferred embodiments.

Examples of Systems and Methods [0063] A method of generating patterns using a network is a step of providing a recursive network of subnetworks having a parent feature node and at least two child feature nodes, the parent feature node. Is an input node and at least two child feature nodes are output nodes; a subnetwork with steps that propagate node selection down the network hierarchy so as to be consistent with the node connections of the subnetwork of the network. Propagation within is the step of selecting one pool node that is consistent with the parent feature node's selection function at the subnetwork's activated parent feature node; at the pool node of the first pool, the subnetwork's child features. With the step of selecting at least the first parent-specific child feature (PSCF) node corresponding to the node; in response to the selection of the first node of the subnetwork, at least the second pool according to the lateral constraint node of the subnetwork. Includes the step of enforcing selection constraints on the second node of the network; the step of compiling the state of the final child feature node of the network into the generated output.

[0064] In one embodiment of this method, the first node is the first PSCF node and the second node in the second pool is the PSCF node. In a further embodiment of this method, the second PSCF node in the second pool is in the subnet of the first node. In a second further embodiment of this method, the second PSCF node in the second pool is in the second subnetwork. In a third further embodiment of this method, the first PSCF node is in the network at the first time and the second PSCF node in the second pool is in the network at the second time. In a further fourth embodiment of this method, the method maps the selected child feature nodes to data features that are compiled into the generated output in response to the selection of the PSCF node that selects at least one child feature node. Includes steps to do.

The generated output of the method may include images, image sequences, and sounds. The choice function of this method can be a Boolean logic choice function.

In the second embodiment of this method, the child feature nodes of the first layer subnetworks are the parent feature nodes of at least two subnetworks of the second layer. In a further embodiment of this method, the step of selecting pool nodes that is consistent with the choice function includes selecting each pool node connected to the parent node according to the logical AND choice function; selecting the first PSCF node. The step includes randomly selecting only one PSCF node from a set of PSCF nodes connected to the pool node according to the pool node's logical XOR selection function. In a second further embodiment of this method, the step of selecting at least the first PSCF node is a first subset of PSCF nodes from a set of PSCF nodes connected to the pool node according to the pool node's logical XOR selection function. This selection function is applied to multiple subsets of a set of PSCF nodes, further including the step of randomly selecting. In a third further embodiment of this method, the step of selecting a pool node that is consistent with the choice function includes selecting each pool node connected to the parent node according to the logical AND selection function.

A method for generating patterns using a network is a step of providing a recursive network of sub-networks with a parent feature node and at least two child feature nodes, where the parent feature node is the input node. Yes, at least two child feature nodes are output nodes; in the first subnetwork of the recursive network, the first subnetwork includes said parent feature node and said at least two child feature nodes. In the activated parent feature node of the first subnetwork, the step of selecting the first pool node and the second pool node that are consistent with the selection function of the parent feature node; in the first pool node. With the step of selecting at least the first parent-specific child feature (PSCF) node corresponding to the first child feature node of the subnetwork; in the second pool node, corresponding to the second child feature node of the subnetwork. A step of selecting at least a second parent-specific child feature (PSCF) node; and a step of compiling the state of the final child feature node of the network, including the first and second child feature nodes, into generated output. And include.

[0068] In one embodiment of this method, the method maps the selected child feature nodes to the data features that are compiled into the generated output in response to the selection of the first and second PSCF nodes. Includes more steps. The generated output of this method may include images, image sequences, sounds.

In a second embodiment of the method, the method provides at least a third parent-specific child feature (PSCF) node in the first pool node that corresponds to a third child feature node in the subnet. The first and third PSCF nodes are selected according to the logical XOR selection function of the first pool node applied over a subset of the set of PSCF nodes connected to the first pool node, including the step of selection. ..

A method for inferring a pattern using a network is a step of providing a recursive network of subnetworks with a parent feature node and at least two child feature nodes; The steps to configure a node; the step to receive data feature input at the final child feature node; and the node activation propagated through the network hierarchy so as to be consistent with the node connections of the network's subnetworks and the posterior prediction of the child nodes It includes a step of executing an activation constraint between at least two nodes of the subnetwork; and a step of outputting the selection of the parent feature node to the inference output.

In one embodiment of this method, the step of propagating node activation is the step of the child feature node messaging the likelihood score to the connected parent-specific child feature (PSCF) node; In the pool node, the step of generating the likelihood score from the likelihood score of the PSCF node connected to the posterior distribution element; and in the parent feature node of the subnetwork, the likelihood score of the parent feature node connected to the posterior distribution component. And from, including a step to generate a likelihood score. In a further embodiment of this method, the step of enforcing the activation constraint between at least two nodes is the first PSCF node connected to the first pool node and the second connected to the second pool node. Includes the step of enforcing activation constraints with the PSCF node. In a further second embodiment of this method, the step of enforcing the activation constraint between at least two nodes is the first PSCF node in the first subnetworks and the second PSCF in the second subnetworks. Includes the step of enforcing activation constraints with the node. In a further embodiment of this method, enforcing an activation constraint between at least two nodes is performed with a first PSCF node in the first time network and a second PSCF in the second time network. Includes steps to enforce activation constraints between.

In a second embodiment of this method, the step of receiving the data feature input includes setting the activation of the final child feature node according to the image feature of the image, and the inferred output is in the image. It is an object detected in.

[0073] In a third embodiment of the method, the step of receiving the data feature input comprises setting the activation of the final child feature node according to the voice signal feature, and the inferred output is the voice signal. The detected voice pattern inside.

[0074] In a fourth embodiment of this method, the child feature nodes of the first layer subnetworks are the parent feature nodes of at least two second layer subnetworks.

[0075] A method of constructing a neural network recursively combines a plurality of subnetworks in a network hierarchy, including communicably joining each child feature node of the upper hierarchy subnetwork to a parent feature node of the lower hierarchy subnetwork. The step to build in; the step to set the selection function of the parent feature node of the subnetwork, the selection function is defined by the selection option of at least two pools in the subnetwork, and; the selection of the pool node. The steps to configure the function, the pool node selection function, is defined by the selection options of at least two parent-specific child feature (PSCF) nodes; and the step of linking at least one pair of nodes with constraint nodes. Includes node connectivity, selection functions, and steps to propagate node selection through the network hierarchy in a manner consistent with linked constraint nodes in the subnetwork of the network.

[0076] In one embodiment of this method, the step of linking at least a pair of nodes with the constraint node is a second step in which the first PSCF node connected to the first pool node is connected to the second pool. Includes a step to link with the PSCF node.

[0077] In a second embodiment of this method, the step of linking at least a pair of nodes with a constraint node is to make the first PSCF node of the first subnet network a second PSCF node of the second subnet network. Includes steps to link. In a further embodiment of the method, the step of linking at least a pair of nodes with a constraint node is to make the first PSCF node in the first network of the first instance the first in the second network of the first instance. Includes a step of linking with 2 PSCF nodes. In a second further embodiment of this method, the method further comprises setting posterior parameters, and the step of propagating node selection through the network hierarchy is further performed in a manner consistent with belief propagation according to the configured posterior parameters. Will be done. In a third further embodiment of the method, the method further comprises connecting a PSCF node to a child feature node, at least one child feature node being connected to at least two PSCF nodes.

[0078] The system comprises a recursively designed network of sub-networks organized in multiple layers, the sub-network being at least a parent feature node, a pool node, and a parent-specific child feature (PSCF) node. And a child feature node, the parent feature node of at least one subnetwork is composed of selection functions that can operate on at least two pool nodes connected to the parent feature node of at least one subnetwork. The pool node of at least one subnetwork consists of select functions that can operate on at least two PSCF nodes connected to the pool node of at least one subnetwork; the PSCF node of at least one subnetwork is connected. It is configured to activate the child feature node that has been created, and the child feature node can connect to at least the parent feature node of the second subnetwork in the lower hierarchy.

[0079] In a first embodiment of the system, the system further comprises a constraint node having at least two connections from at least two PSCF nodes and having a choice function to enhance selection by the pool node. .. In a further embodiment of the system, the constraint node is connected to the first PCSF node and the second PCSF node, the first PCSF node is connected to the first pool node, and the second PCSF node is the second pool. Connected to the node. In a further second embodiment of the system, the first pool node and the second pool node are connected to a single shared parent feature node, and the first pool node and the second node are in a single subnet. include. In a further third embodiment of the system, the first pool node is connected to the first parent feature node and is contained within the first subnetworks; the second pool node is connected to the second parent feature node. It is included in a second subnetwork different from the first subnetwork. In a further fourth embodiment of the system, the first subnetworks are at a different hierarchy level than the second subnetworks.

[0080] In a second embodiment of this system, the nodes of the network are composed of posterior distribution elements; the network is connected through nodes with subnetworks of the network and through multiple layers in a manner consistent with the posterior distribution elements. Configured to propagate node activation. In a further embodiment of the system, the network is configured to output the selection of the parent feature node as an estimated output in response to receiving the data feature input on at least one child feature node. In a further second embodiment of the system, the network: at the child feature node, with the step of messaging the likelihood score to the connected parent-specific child feature (PSCF) node; at the pool node of the subnetwork, the posterior distribution element. And the step of generating the likelihood score from the likelihood score of the PSCF node connected to; in the parent feature node of the subnetwork, the likelihood score from the posterior distribution element and the pool node connected to the parent feature node. It is configured to propagate node activation by steps to generate a degree score. In a further third embodiment of the system, the network is configured to receive data feature inputs by setting the activation of child feature nodes according to the image features of the image, where the inference output includes image classification. .. In a further fourth embodiment of the system, the network is configured to receive data feature inputs by setting the activation of child feature nodes according to the voice features of the voice signal, where the inference output classifies the voice signal. Including.

[0081] In a third embodiment of the system, the network is configured to output a child feature node selection as a generated output in response to the received parent feature input. In a further embodiment of the system, the generated output is an image generated based on the mapping of selected child feature nodes to image features. In a further second embodiment of the system, the generated output is an image generated based on the mapping of selected child feature nodes to image features. In a further third embodiment of the system, the parent feature input is image classification. In a further fourth embodiment of the system, the generated output is a voice signal generated based on the mapping of selected child feature nodes to voice features. In a further fifth embodiment of the system, the parent feature input is voice signal classification. In a sixth further embodiment of the system, the network is further configured to receive seed child feature inputs, where the seed child feature inputs are used to modify the output produced by the network. In a further seventh embodiment of the system, the network is configured to receive seed child feature inputs by configuring activation of seeded child feature nodes. In a further eighth embodiment of the system, the network is configured to output the child feature node selection as a generated output in response to both the received parent feature input and the seed child feature input. In a further ninth embodiment of the system, the network is configured to set the activation of the seeded child feature nodes according to the image features of the first part of the image, and the generated output is the second part of the image. ..

In a fourth embodiment of the system, the child feature nodes are connected to a plurality of parent feature nodes in the subnetwork of the lower hierarchy. In a further embodiment of this system, the evidence provided to the child feature nodes is the evidence transmitted by the first parent feature node of the plurality of parent feature nodes and the first of the plurality of parent feature nodes. It is processed by applying a noisy-OR gate to the evidence transmitted by the parent feature node of 2. In a further second embodiment of the system, the evidence provided to the child feature nodes is the evidence sent by the first parent feature node of the plurality of parent feature nodes and the first of the plurality of parent feature nodes. It is processed by applying a noisy-MAX gate to the evidence transmitted by the parent feature node of 2.

Claims (56)

A method of generating patterns using a network,
A step of providing a recursive network of subnetworks having one parent feature node and at least two child feature nodes, wherein the parent feature node is an input node and the at least two child feature nodes are output nodes. There are steps and;
The step of propagating node selection down the network hierarchy based on the node connections of the subnetwork of the network, the propagation within the subnetwork is:
In the activated parent feature node of one subnetwork, the process of selecting one pool node using the selection function of the parent feature node;
In the pool node of the first pool, the step of selecting at least the first parent-specific child feature (PSCF) node corresponding to the child feature node of the subnetwork;
A step comprising executing a selection constraint on at least a second PSCF node in the second pool according to the lateral constraint node of the subnet in response to the selection of the first PSCF node ;
A method comprising: compiling a selection of the final child feature nodes of the network into generated output. The method according to claim 1 , wherein the second PSCF node of the second pool is in a subnetwork of the first PSCF node. The method according to claim 1 , wherein the second PSCF node of the second pool is in the second subnetwork. The method of claim 1 is characterized in that the first PSCF node is in the network at the first time and the second PSCF node in the second pool is in the network at the second time. Method. In the method of claim 1 , in response to the step of selecting the PSCF node, a step of selecting at least one child feature node and mapping the selected child feature node to a data feature compiled into the generated output. A method characterized by further including steps to be performed. The method according to claim 5 , wherein the generated output is an image. The method according to claim 5 , wherein the generated output is an image sequence. The method according to claim 5 , wherein the generated output is voice. The method according to claim 5 , wherein the selection function is a binary logic selection function. The method according to claim 9 , wherein the child feature node of the first layer subnetwork is the parent feature node of at least two subnetworks of the second layer. In the method of claim 10 , the step of selecting a pool node that is consistent with the selection function includes selecting each pool node connected to the parent feature node according to the logical AND selection function, and the first PSCF node is selected. The method of selecting includes a step of randomly selecting only one PSCF node from a set of PSCF nodes connected to the pool node according to the logical XOR selection function of the pool node. In the method of claim 10 , the step of selecting at least the first PSCF node randomizes the first subset of PSCF nodes from the set of PSCF nodes connected to the pool node according to the logical XOR selection function of the pool node. A method comprising further selecting in, wherein the selection function is applied to a plurality of subsets of the set of PSCF nodes. In the method of claim 12 , the step of selecting a pool node consistent with the selection function includes a step of selecting each pool node connected to the parent feature node according to the logical AND selection function. A method of generating patterns using a network,
A step of providing a recursive network of subnetworks having a parent feature node and at least two child feature nodes, wherein the parent feature node is an input node and the at least two child feature nodes are output nodes. With steps;
In the first subnetwork of the recursive network, the first subnetwork includes said parent feature node and said at least two child feature nodes.
In the activated parent feature node of the first subnetwork, the step of selecting the first pool node and the second pool node that are consistent with the selection function of the parent feature node;
With the step of selecting at least the first parent-specific child feature (PSCF) node corresponding to the first child feature node of the subnet in the first pool node;
With the step of selecting at least a second parent-specific child feature (PSCF) node corresponding to the second child feature node of the subnet in the second pool node;
A method comprising: compiling a node selection of the final child feature node of the network into a generated output, including said first and second child feature nodes. The method of claim 14 further comprises the step of mapping the selected child feature nodes to the data features compiled into the generated output in response to selecting the first and second PSCF nodes. A method characterized by. A method according to claim 15 , wherein the generated output is an image. A method according to claim 15 , wherein the generated output is an image sequence. The method according to claim 15 , wherein the generated output is voice. The method of claim 14 further comprises the step of selecting at least a third parent-specific child feature (PSCF) node in the first pool node that corresponds to the third child feature node of the subnet. The first and third PSCF nodes are selected according to the logical XOR selection function of the first pool node applied across a plurality of subsets of the set of PSCF nodes connected to the first pool node. How to feature. A method for inferring patterns using a network,
With the step of providing a recursive network of subnetworks with a parent feature node and at least two child feature nodes;
With the steps to configure the nodes of the subnetwork using the posterior distribution model ;
With the step of receiving the data feature input at the final child feature node;
With the steps of propagating node activations through the network hierarchy in a way that is consistent with the node connections of the network's subnetworks and the posterior prediction of child nodes;
With the step of enforcing activation constraints between at least two nodes in the subnetwork;
A method characterized by including a step of outputting the parent feature node selection to the inference output. In the method of claim 20 , the step of propagating node activation is the step of the child feature node sending a likelihood score to the connected parent-specific child feature (PSCF) node;
In the pool node of the subnetwork, the step of generating the likelihood score from the posterior distribution model and the likelihood score of the connected PSCF node;
A method comprising a step of generating a likelihood score from a posterior distribution model and a likelihood score of a pool node connected to the parent feature node in a parent feature node of the subnet. In the method of claim 21 , the step of enforcing the activation constraint between at least two nodes is a first PSCF node connected to a first pool node and a second connected to a second pool node. A method comprising the step of enforcing an activation constraint with a PSCF node. In the method of claim 21 , the step of enforcing the activation constraint between at least two nodes is the first PSCF node in the first subnet network and the second PSCF node in the second subnet network. A method characterized by including the step of enforcing activation constraints between. In the method of claim 21 , the step of enforcing the activation constraint between at least two nodes includes a first PSCF node in the network at the first time and a second PSCF in the network at the second time. A method characterized by including a step of enforcing an activation constraint between. In the method of claim 20 , receiving the data feature input comprises setting the activation of the final child feature node according to the image feature of the image, the inference output being the object detected in the image. A method characterized by that. In the method of claim 20 , receiving the data feature input comprises setting the activation of the final child feature node according to the voice signal feature, said inference output being the detected voice pattern in the voice signal. A method characterized by being. The method according to claim 20 , wherein the child feature node of the first layer subnetwork is a parent feature node of at least two second layer subnetworks. In a method of constructing a neural network implemented by a processor, the processor
Steps to recursively build multiple subnetworks in the network hierarchy, including communicatively joining each child feature node of the upper layer subnetworks to the parent feature node of the lower hierarchy subnetwork;
The step of setting the selection function of the parent feature node of the subnetwork, wherein the selection function is defined by the selection options of at least two pools in the subnetwork.
Comprising the steps of setting the selecting function of the pool node, select functions of the pool node is defined by at least two parental specific child characteristics (PSCF) node selection options, the steps;
With the step of linking at least a pair of nodes with the constraint node;
A method comprising a node connection of a sub-network of the network, the selection function, and a step of propagating the node selection through the network hierarchy in a manner consistent with the linked constraint node. In the method of claim 28 , the step of linking at least a pair of nodes with the constraint node is to connect the first PSCF node connected to the first pool node to the second PSCF node connected to the second pool. A method characterized by including a step of linking. In the method of claim 28 , the step of linking at least a pair of nodes with the constraint node comprises linking the first PSCF node of the first subnet network with the second PSCF node of the second subnet network. A method characterized by that. In the method of claim 28 , the step of linking at least a pair of nodes with the constraint node is to make the first PSCF node of the first network in the first instance a second of the second network in the first instance. A method comprising linking with a PSCF node. The method of claim 28 , further comprising setting a posterior parameter, characterized in that the step of propagating the node selection through the network hierarchy is further executed in a manner consistent with belief propagation according to the set posterior parameter. how to. A method according to claim 28 , further comprising connecting the PSCF node to a child feature node, wherein at least one child feature node is connected to at least two PSCF nodes. It's a system
It has a recursively designed network of sub-networks organized in multiple layers.
The subnetworks include at least a parent feature node, a pool node, a parent-specific child feature (PSCF) node, and a child feature node;
The parent feature node of the at least one subnetworks consists of a choice function that selects one pool node from at least two pool nodes connected to the parent feature node of the at least one subnetworks;
The pool node of the at least one subnetwork is composed of a selection function that selects one PSCF node from at least two PSCF nodes connected to the pool node of the at least one subnetwork;
PSCF nodes in at least one subnetwork are configured to activate connected child feature nodes;
A system characterized in that the child feature node can be connected to at least a parent feature node of a second subnetwork in a lower hierarchy. The system of claim 34 further comprises a constraint node having at least two connections from at least two PSCF nodes and having a third choice function for enhancing selection by said pool node. .. In the system of claim 35 , the constraint node is connected to a first PCSF node and a second PCSF node, the first PCSF node is connected to a first pool node, and the second PCSF node is A system characterized by being connected to a second pool node. In the system of claim 36 , the first pool node and the second pool node are connected to a single shared parent feature node, and the first pool node and the second node are single subs. A system characterized by being contained within a network. In the system of claim 36 , the first pool node is connected to the first parent feature node and is included in the first subnetworks, and the second pool node is the second parent feature. A system characterized in that it is connected to a node and is included in a second subnetwork different from the first subnetwork. The system according to claim 38 , wherein the first sub-network is at a different hierarchical level than the second sub-network. In the system of claim 34 , the nodes of the network are configured with a posterior distribution model , and the network is node-acquired through multiple layers in a manner consistent with the node connections of the subnets of the network and the posterior distribution model. A system characterized by being configured to propagate a titation. The system of claim 40 is characterized in that the network is configured to output a selection of parent feature nodes as inference output in response to receiving data feature inputs at at least one child feature node. System to do. In the system of claim 41 , the network is:
• In the child feature node, the step of messaging the likelihood score to the connected parent-specific child feature (PSCF) node;
-In the pool node of the subnetwork, the step of generating the likelihood score from the posterior distribution model and the likelihood score of the connected PSCF node:
A step of generating a likelihood score from the posterior distribution model and the likelihood score of the pool node connected to the parent feature node in the parent feature node of the subnetwork.
A system characterized by being configured to propagate node activation by. In the system of claim 41 , the network is configured to receive data feature inputs by setting the activation of child feature nodes according to the image features of the image, and the inference output comprises the classification of the image. System. In the system of claim 41 , the network is configured to receive data feature inputs by setting the activation of child feature nodes according to the voice features of the voice signal, and the inference output comprises a classification of the voice signal. Characteristic system. The system according to claim 40 , wherein the network is configured to output a child feature node selection as a generated output in response to a received parent feature input. The system according to claim 45 , wherein the generated output is an image generated based on a mapping of selected child feature nodes to image features. The system according to claim 46 , wherein the parent feature input is an image classification. The system according to claim 45 , wherein the generated output is a voice signal generated based on a mapping of the selected child feature node to the voice feature. The system according to claim 46 , wherein the parent feature input is a voice signal classification. In the system of claim 45 , the network is further configured to receive a seed child feature input, which seed child feature input is used by the network to modify the generated output. system. The system of claim 50 , wherein the network is configured to receive a seeded child feature input by setting activation of the seeded child feature node. The system of claim 51 is characterized in that the network is configured to output a child feature node selection as a generated output in response to both the received parent feature input and the seed child feature input. System to do. In the system of claim 52 , the network is configured to set the activation of the seeded child feature node according to the image features of the first portion of the image, the generated output being the second of the images. A system characterized by being a part. The system according to claim 34 , wherein the child feature node is connected to a plurality of parent feature nodes of a subnetwork at a lower hierarchy level. In the system of claim 54 , the evidence provided to the child feature node is the evidence transmitted by the first parent feature node of the plurality of parent feature nodes and the second of the plurality of parent feature nodes. A system characterized in that it is processed by applying a Noisy-OR gate to the evidence transmitted by the parent feature node of. In the system of claim 54 , the evidence provided to the child feature node is the evidence transmitted by the first parent feature node of the plurality of parent feature nodes and the first of the plurality of parent feature nodes. A system characterized in that the evidence transmitted by the parent feature node of 2 is processed by applying a Nosy-MAX gate.

Images