JP6761055B2 - Systems and methods for generating data interpretations in neural networks and related systems - Google Patents

Description

[0001] The present invention relates generally to the field of artificial intelligence, and more specifically to novel and useful systems and methods for generating data interpretations for neural networks, and related systems 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. Models capable of performing visual object recognition must be able to provide a description of the visual dataset in order to recognize the objects that exist within those visual datasets. Visual object recognition is a particular case of pattern recognition (and vice versa), a more common problem in artificial intelligence. Pattern recognition is a problem in other disciplines and media other than image processing, such as speech recognition, natural language processing, and other disciplines. Therefore, there is a need in the field of artificial intelligence to create new and useful methods for generating data interpretations for neural networks and related systems. The present invention provides such a novel and useful method.

FIG. 1 is a flowchart of a layer-based bidirectional data conversion system (LBD). FIG. 2 is a schematic diagram of a recursive cortical network (RCN). FIG. 3 is a schematic diagram of a recursive cortical network (RCN). FIG. 4 is a schematic diagram of the RCN subnetworks. FIG. 5 is a schematic diagram of the RCN subnetworks. FIG. 6 is a chart of the method of the preferred embodiment. FIG. 7 is a schematic diagram of inference using forward transformation. FIG. 8 is a schematic diagram of inference using a combination of forward and reverse transformations. FIG. 9 is a flowchart of forward conversion of the method of the preferred embodiment. FIG. 10 is an exemplary diagram of the LBD. FIG. 11 is a diagram showing an exemplary implementation example of forward conversion of the method of the preferred embodiment. FIG. 12 is a diagram showing an exemplary implementation example of forward conversion of the method of the preferred embodiment. FIG. 13 is a diagram showing an exemplary implementation example of the reverse conversion of the method of the preferred embodiment. FIG. 14 is a flowchart implementation example of the reverse conversion of the method of the preferred embodiment. FIG. 15 is an exemplary implementation of the reverse transformation of the method of the preferred embodiment.

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.

Systems and methods for generating data interpretations of preferred embodiments function to improve the generation and / or inference tasks of neural networks and related systems. This system and method preferably applies bidirectional transformation through various layers of a data transformation system (eg, a neural network). To address the challenges of pattern recognition, the systems and methods of preferred embodiments can be applied to generate data explanations for pattern data. The generation of data interpretations is important for many pattern recognition models, including convolutional neural networks, recursive cortical networks, and other models consisting of a series of layers, each layer applying transformations to the underlying layers. In the first objective task of the neural network, an inference output is generated. The process and system for producing such inference output is described herein. In particular, in one example, the inference output can be preferably improved by applying a generational transformation in the opposite direction. A second objective task of a neural network can provide a generated output or an imagining output. The systems and methods described herein may, in addition or alternative, be applied to provide such generated output. In one variant of the production process, it is preferable to use inference transformations, at least in part, to preferably improve the production output. This system and method is preferably used in neural networks, more specifically recursive cortical networks, but this system and method can also be used in any suitable layered data conversion system.

1. 1. Neural Networks and Related Systems [0020] Neural networks and related systems include recursive cortical networks (RCN), convolutional neural networks (CNN), HMAX models, slow feature analysis (SFA) systems, and hierarchical time storage (HTM) systems. Can be used for a wide variety of tasks that are difficult to complete using standard rule-based programming, including. These tasks include many of the key areas of computer vision and speech recognition.

Neural networks and related systems can be represented as distributed processing elements that perform addition, multiplication, exponentiation or other functions on the element's incoming message / signal. Such networks can be implemented and implemented through various implementation examples. For example, a system can be implemented as a network of electronically coupled functional node components. These functional node components can be logic gates that are placed or configured within the processor to perform a particular function. As a second example, the system can be implemented as a network model programmed and / or configured to run on a processor. Such a network model is preferably electronically stored software that encodes the operation and communication between the nodes of the network. Neural networks and related systems can be used in a wide variety of applications, including images, videos, audio, natural language text, analytical data, widely distributed sensor data, or other suitable data formats. Data types can be used.

Neural networks and related systems can be described as systems that include multiple transformation layers. The input data (usually having a low level of abstraction) entering these systems is transformed in the first layer to produce the first intermediate data. The first intermediate data is converted in the second layer to generate the second intermediate data, and so on. This process continues until the system reaches the final layer, where the intermediate data produces output data (usually with a higher level of abstraction). This process produces a data interpretation (ie, inference) for a dataset, such as by identifying local features of the dataset and identifying more complex features based on these local features. Used to increase the level of abstraction of data interpretation at each layer.

[0023] One layer may generate both intermediate data and output data. That is, the output of one intermediate layer can be used both as an input to a higher layer and as a data interpretation output (ie, an output to another process).

Several neural networks and related systems can be used in a complementary manner. In this way, a system similar to the one described above can be initialized from the final layer (or intermediate layer) to transform the data from a higher level of abstraction to a lower level of abstraction. This process is commonly referred to as generation (and if generation is not affected by the data generated from the input data, it is called imagination).

Especially for artificial intelligence applications (eg computer vision), neural networks and related systems can perform both inference and generation by redirecting data propagation, as shown in FIG. In such a system, hereinafter referred to as a layer-based bidirectional data conversion system (LBD), inference is performed by passing a "bottom-up message" (for example, BU1, BU2 ...) from a lower layer to an upper layer. Will be executed. On the other hand, the generation is executed by passing a "top-down message" (for example, TD1, TD2 ...) From the upper layer of the system to the lower layer of the system. In LBD, the number of layers may be arbitrarily large to allow for various gradations of the input / output conversion process.

As an example of inference, the input data to the LBD may consist of an image. This image data may be introduced directly into the LBD (as a BU1 message) or preprocessed (eg, by increasing the image contrast of the input data for transmission as a BU1 message). A BU1 message consisting of image data (preprocessed or unprocessed) is transformed by layer 1 to create a converted dataset (eg, a contour detected by position) represented as a BU2 message. This BU2 message is further converted by layers 3 and 4, and BU3 and BU4 messages are sequentially created, respectively. The BU4 message in this case represents, for example, the dataset corresponding to the detected object in the image data. This BU4 message may be processed by the post-processing layer (eg, processing BU4 into a description that describes the objects in the image).

As a similar example of generation, the LBD may have a generation constraint that has passed through the post-processing layer (eg, a description that describes an object that is desired to be in the generated image), or more directly a TD1 message. (For example, a dataset corresponding to an object that is desired to be included in the generated image data) is given. The TD1 message is converted into a TD2 message by layer 3 and then converted into a TD3 message by layer 2. The transformation of these layers effectively predicts possible lower level of abstraction data based on the inputs to the layers (eg, TD1 message in the case of layer 3). Finally, the LBD can output generated data that represents the image data predicted by the system based on generation constraints or other initialization inputs. Additional or alternative, the LBD is not given input data for generation, resulting in an assumption of default values (or randomly generated values) stored in the system. The output of this LBD is a special case of generation called imagination.

The previous two examples represent inference of input data and / or BU1 messages and generation based on generation constraints and / or TD1 messages, but such inferences or generations can occur at any layer of the LBD. .. For example, the LBD may perform inferences on the data provided directly to layer 3. Neural networks and related systems preferably apply complementary generation and / or inference during the inference and / or inference process.

2. 2. Recursive Cortical Network [0029] The systems and methods of preferred embodiments described herein are preferably applicable to any neural network and related system (ie, LBD) that fits the above description, but implementation details. And examples are particularly relevant to the retrocortical network (RCN).

As shown in FIG. 2, the recursive cortical network (RCN) may include multiple subnetworks. The subnet of 1 preferably includes at least a parent feature node, a pool node, a parent-specific child feature node (or PSCF node for short), and at least one constraint node. The RCN can be configured for different modes of operation, including a first mode of operation: generation mode and a second mode: inference mode. As shown in FIG. 3, the RCN is preferably a hierarchically organized network of interconnected subnetworks having various parent-child relationships. Alternatively, the RCN can be a single layer or a single subnet of a collection of subnetworks.

As shown in FIG. 4, various instances and instantiations of the RCN subnetworks are recursively constructed, connected, and used in the RCN hierarchy. Hierarchical network architectures can be built algorithmically, or at least through partial user selection and configuration. RCN 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. 4, 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 an input in generation and an output in inference. Conversely, child node connections are outputs in generation and inputs in inference. In a variant of a single-layer (or non-layer) subnetworks, the subnetworks are placed as siblings. Subnetworks such as those described below can interact through various forms of constraint nodes.

Subnetworks can be set up in a variety of different configurations within the network. Many configurations are determined by constrained nodes that define node selection within, between subnetworks, or even between networks. 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, a parent feature node of a subnetwork can participate as a child feature node of a higher subnetwork. 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 multiple subnetworks, multiple parental interactions can be introduced at several nodes in the network. These interactions can be modeled using different probabilistic models within the node.

Hierarchical connection of subnetworks serves to facilitate compact and compressed representation through the reuse of subnetworks. 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 representations of child subnetworks can be reused across multiple parent subnetworks. One example where this is applicable is when the RCN represents a visual object. Lower subnetworks can correspond to parts of an object, and higher subnetworks (ie, upper subnetworks) can represent how those parts come together to form an object. For example, lower level subnetworks can accommodate the representation 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 RCN can be used both to generate data interpretations (eg, classify objects in an image) and to generate data predictions (eg, images containing a set of objects). When generating a data interpretation, preferably the nodes of the RCN act on the input data features and propagate the node selection / processing through the hierarchy of the RCN until the output from the parent feature of the top-level subnetworks is obtained. To achieve this output, a combination of propagating information up (to the upper parent layer) and down (to the final child feature) of the hierarchy can be used. When generating data predictions, the RCN preferably starts with a general generated request directed to, supplied, or delivered to the parent feature node of the top-tier subnetworks. It is preferred that the nodes act on the information and propagate the node selection / processing down the RCN hierarchy until output is obtained from the child feature nodes of the bottom subnetwork. As shown in FIG. 5, the sub-network functions to provide a node selection operation between the parent feature and the child feature. Subnetworks are a basic component of RCN. Subnetworks, in the case of generation, are mapped or networked from one higher level feature to a set of lower level features, whereby the activity of the lower level features (eg, the visual features of the image). ) Is determined by a higher level feature activity (eg, object name). In the case of inference, the subnetworks are mapped or networked from lower level features to higher level features so that the activity of the higher level features (eg object name) is the activity of the lower level features. Determined by (eg, visual features of the image). The general architecture of a subnetwork preferably includes a single top-level node that is the parent feature node. The parent feature node (PF1) preferably includes connections to at least two pool nodes (P1 and P2). Each pool node preferably includes connections to a plurality of PSCF nodes 130 (X1, X2, X3, X4, X5, X6). Constraint nodes (C1, C2, C3) may also be in the subnetwork. The constraint node is preferably connected to another PSCF node. Constraint nodes define restrictions, rules, and restrictions between at least two PSCF nodes. The PSCF node is preferably connected to the child feature nodes 150 (CF1, CF2, CF3, CR4, CF5, CF6). Subnetwork instances within the RCN 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 (eg, parent feature node, child feature node) is a binary random variable node, preferably having a plurality of parents and a plurality of children. When having multiple parents (ie, multiple nodes are connected via a parent / input connection), the interaction between the parent connections is preferably treated as a superposition of connections. Additional or alternative, multi-parent interactions can be modeled in any way. Multi-parent interactions can be stochastically modeled within a node using canonical models such as Nosy-OR and Nosy-MAX gates. The child connection of the feature node preferably encodes the stochastic relationship between the feature and the pool. In some RCNs, if one feature is active, then all pools of that feature are active, but such activations can be modified according to probability tables 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).

When the pool node is on, p and q are zero only when the feature is on. However, other values of p and q may be used instead. The pool node is preferably treated as a binary node. The pool node preferably has one parent connection representing the probability table described above. Pool nodes can have multiple connections to 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 relevant probability. In other words, the instantaneous connection represents a multinomial random variable connection. Pool elements (modeled as pool members, a possible activation set of PSCF nodes) are preferably configured to act as binary random variables, at least one of which is pooled according to the distribution P (M | pool). Is selected when is selected. The pool element represents a combination of child feature functions. 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 constraint is often an AND or OR constraint, but it may be any suitable choice function. These constrained nodes can also be nodes with connections larger than the paired connections.

[0037] The parent feature node functions as a higher-level feature node. In the generation operation mode, the parent feature node is the input of the subnetwork. In the inference operation mode, the parent feature node is the output of the subnetwork. The parent feature node is configured to perform a choice function upon activation. The choice function is preferably a logical function such as an AND, OR, NOT, or Boolean choice 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. In addition, the randomized selection may be biased or weighted according to the node connection weighting of the connection between the parent feature node and the pool node. The choice function can be, as an alternative, a probability choice function or any suitable function used in selecting connection options.

The pool node functions as a node for selecting from a set of child features. The child features associated with one pool node 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, PSCF nodes are preferably an invariant representation of a variation of features. In FIG. 5, P1 is an invariant representation of 1 for 3 different translations of the vertical line and P2 is an invariant representation of 1 for 3 different transformations 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. A possible set of PSCF nodes is preferably any PSCF node that has a connection to a pool node. The pool may be constrained. For example, the elements of the pool 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, the pool node is configured to execute a choice function at activation. This choice function can be any suitable function, but is preferably a logical operator as described above for the parent feature node. Choice functions can be randomized, biased and / or weighted as well. The pool node selection function preferably signals the corresponding PSCF node by selecting, triggering, activating, or otherwise. In addition, the choice function may be restricted or disabled based on the activated constraint node. Activated constraint nodes may define which nodes are selected in the pool based on the selection of PSCF nodes (connected through the constraint nodes). Similarly, it can determine the possible set of PSCF nodes for the pool node and / or the weighting or preference of the pool node. Pool nodes in a subnetwork can be evaluated sequentially so that constraint nodes can be applied to other pools as needed.

The PSCF node serves as an option for the immutable feature option. One PSCF node maps to one child feature and one PSCF node has only one parent pool node. PSCF nodes may be further connected or coupled with constraint nodes. The constraint node preferably defines the relationship between the plurality of PSCF nodes. The constraint node preferably connects to other PSCF nodes in different pools, different times, and / or different subnetworks. PSCF nodes are preferably not shared between subnetworks. However, child feature nodes (which can be parent nodes of subnetworks) may share connections to multiple subnetworks.

The constraint node functions to limit the types of patterns allowed in the subnetwork. The constraint node is preferably connected to at least two PSCF nodes. Alternatively, more than 2 PSCF nodes may be connected via 1 constraint node. Constraint nodes may also be between any suitable type of node. Constraint nodes may be between pool nodes. The constraint node may also be between two types of nodes. For example, a constraint node can connect a PSCF node and a pool node. Here, as a preferred implementation example, constraint nodes connect PSCF nodes to each other, but constraint nodes are used to impose constraints between any set of nodes (of any kind) in the RCN. can do. Constraint nodes can be between pool nodes, between pool nodes and PSCF nodes, or between any suitable node in the network. The PSCF nodes are preferably not in the same pool and in some cases are not in the same subnetworks. The constraint nodes preferably connect PSCF nodes of the same layer to each other, but may instead connect subnetworks of different layers. Furthermore, one constraint node may be connected to any suitable PSCF node, or any suitable number of constraint nodes may be connected. Constraint nodes can enforce restrictions, rules, and constraints within the selection of other pools, other subnetworks, and / or nodes at different times. The network is preferably evaluated in an ordered manner so that PSCF nodes connected via constraint nodes are not preferably evaluated at the same time. When the first PSCF node is activated or selected, the constraint node connected to the first PSCF node is activated. The restriction of the constraint node is then activated / enforced on the connected PSCF node. The constraint node, like any other node, may have a choice function that determines how to activate the PSCF node. The constraint node preferably affects how the pool node can select the PSCF node. In one example, the constraint node choice function is an AND logical operator that performs a selection of connected PSCF nodes when one of the PSCF nodes is active. In another variant, the constraint node choice function can be an OR logical operator that modifies possible PSCF nodes in the pool. You can use any suitable choice function. Some constraint nodes may have a basic or simple constraint that the 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.

[0041] The constraint node can include a lateral constraint node, an external constraint node, and a time constraint node. Lateral constraint nodes serve to limit the pattern of subnet types based on the interaction between the subnet's pool nodes. External constraint nodes serve to enforce immutable patterns across different subnetworks. Similar to how laterally constrained nodes ensure that representations of different pools are consistent with each other by imposing constraints on which PSCF nodes in one pool node can be combined with PSCF nodes in another pool. External constraint nodes can maintain the suitability of the entire hierarchy. Time-constrained nodes serve to enforce relationships across RCNs and subnetworks running on other time instances. At the base level, pool elements (eg, PSCF nodes with shared parent pool nodes) can have relationships that specify the order in which they occur in time. A time-constrained node is preferably a simple direct connection constraint, where the activation / selection of one node enforces the selection of a particular node in the second instance. In another description, constraint nodes behave similarly to specifications in Markov chains.

Two or more types of constraint nodes can be executed on the PSCF node. Lateral constraint nodes impose coordination between PSCF nodes in different pools of the same network, and external constraint nodes impose coordination between PSCF nodes in different subnetworks. Constraint nodes are preferably configured so that conflicts do not occur (eg, if one constraint activates the node and the other specifies that it should not be activated). Conflicts and conflicts between constraint nodes can be resolved using techniques for ranking constraint nodes, or the order in which constraint nodes are enforced, or other appropriate rules.

3. 3. Data Interpretation Generation Method [0043] As shown in FIG. 6, the method 100 for generating the data interpretation in the layer-based bidirectional data conversion system (LBD) includes a step (S110) of receiving evidence data and a conversion configuration. It includes a step of setting (S120), a step of executing forward conversion (S130), a step of executing reverse conversion (S140), and a step of outputting converted evidence data (S150).

Method 100 functions to generate an interpretation of the evidence data received by the layer-based bidirectional data conversion system (LBD). Method 100 is used to infer patterns in 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, etc. Can be done. In one implementation example, method 100 can be used for image detection to detect the presence of an object in an image or video. Method 100 may be used to classify the detected objects, either additionally or alternatively.

[0045] The method 100 finally outputs the converted evidence data (received in step S110) through a series of forward conversion and reverse conversion (steps S130 and S140, respectively) (step S150). , Produces an interpretation, which can be understood or used as an interpretation of the received evidence data. Forward and reverse transformations can be performed on an entire set of evidence data or a subset of evidence data, and the set of transformed evidence data may vary from transformation to transformation. In addition, forward and reverse transformations can be performed in any order and at any time, including simultaneous. Details regarding the selection, ordering, and timing of the conversion dataset are preferably indicated by the conversion configuration (set in step S120).

In general, the forward transformation can be thought of as providing an interpretation of the evidence data, and the reverse transformation can be thought of as predicting the evidence data given a particular interpretation. As mentioned above, forward and reverse transformations can also be thought of as increasing or decreasing the level of abstraction of a particular piece of data. Although forward and reverse transformations act as directions of action through the layers of the LBD, the method of preferred embodiments preferably applies both forms of transformation to improve output. These descriptions are intended as a guide to understanding forward and reverse transformations and are intended to limit or define forward and reverse transformations (discussed in more detail herein). It's not something to do.

Inverse transformations behave substantially opposite to forward transformations in the direction of abstraction (eg, forward transformations and inverse transformations act as abstraction-level increase and decrease, respectively. ), The method can preferably assist in applying a reverse transformation to increase the level of abstraction of the data (eg, generate an interpretation of the evidence data). Consider a reference unidirectional neural network designed to recognize characters in an input image, as shown in FIG. An example of this reference neural network consists of multiple subnetworks (Sa, Sb, Sc) designed to recognize characters in different sections (a, b, c) of the image. The network then outputs the characters detected by these subnetworks in the order corresponding to the order of the image sections (ie, a → b → c) at the higher level nodes. The neural network is designed so that the subnet networks Sa, Sb, and Sc transform the input data into intermediate data that includes a likelihood distribution across characters and location space. For example, one element of the intermediate data output by Sa includes the likelihood that the character "a" is present in the position vector (x1, y1) given the input data. This intermediate data is processed by the sub-network O to generate a likelihood distribution of the string across the string and the position space, which is output as the most probable string. In the case of the image example of FIG. 2, this most likely character string is "cco", which is clearly not the character string ("do") shown in the image. In this example of neural network processing, the neural network of FIG. 7 returned erroneous results because it could not take into account the circumstances surrounding each image section. This can be addressed by attempting to incorporate contextual knowledge into layer 2, for example storing prior probabilities of character combinations given to characters in adjacent image sections. For example, there is some possibility that the "c"? "L" pair is actually a "d", or there may be various strings (eg, "cco" in a particular language). It may be less likely). This approach becomes difficult if the likelihood of different character pairs is equally high, or if the knowledge stored in the neural network is inadequate (eg, cl? L pairs and "d" possibilities). Is equally high, or the relative likelihood of "cco" and "do" in the training set does not represent real-world data). Another approach is to include more data from layer 1, for example, instead of sending character and position data, layer 1 may also include character feature and position data. This approach is, in a sense, similar to simply performing feature detection on an image (rather than an image section) and is problematic because it involves a significant increase in the data transmitted from layer 1 to layer 2.

Applying bidirectional processing to this method preferably addresses the problem of the above reference neural network example and serves as an example of generation to improve inference. The neural network in the above reference example can be similar in all respects, except that the processing of the neural network is bidirectional, applying both forward and reverse transformations. This neural network can utilize recursion to allow context (eg, higher level data) to affect the output of lower level subnetworks, as shown in FIG. .. In the first step of recursion, layer 1 outputs the same likelihood distribution as in a unidirectional network. Instead of outputting the string "cco", layer 2 passes information about the neighborhood area to each subnet network during the second step (eg, Sa information about the contents of Sb). This information can be the total likelihood distribution or a subset thereof (eg, the top 5 likelihoods), or other suitable contextual information. Subnetworks process this information and Sb calculates a substantially modified likelihood distribution (in particular, the likelihood of including "d" and "o" is higher than the previous maximum selection value "k". ). As a result, in step 3, the likelihood distribution with updated Sb is transmitted to O. Then, in step 4, O sends the updated likelihood information to Sa and Sc, which then recalculate their likelihood based on the updated likelihood information, which in step 5 Is returned to O. This process can be repeated as many times as needed until the threshold of step N (eg, number of recursive steps, maximum threshold likelihood, etc.) is reached, at which point O is the final character. Output the column. The recursive steps of this exemplary neural network are similar to the possible forward and reverse transformations of Method 100. Note that in more complex neural networks (or other LBDs), recursive steps can include multiple forward transforms before the reverse transform, and vice versa. The order and direction of conversion is preferably determined by method 100. Bidirectional processing can be applied for the ultimate purpose of generating inference output or for the purpose of generating generated output. Method 100 is preferably performed by a recursive cortical network as described above, by a recursive cortical network as described in US Patent Application No. 13/895225, which is incorporated herein by reference in its entirety. It is more preferred to be carried out. Additional or alternative, method 100 can be performed by any suitable neural network or associated system (ie LBD).

The step S110 of receiving the evidence data functions to provide the LBD with the data for which data interpretation is desired. The step of receiving evidence data includes, but additionally, the step of receiving preprocessed, more likely data features (eg, attribute specifications and related values), reduced, transformed, or extracted data. Alternatively, it may include unprocessed data. For example, the image may be subdivided into a plurality of image blocks, and the pixel patterns in the plurality of blocks may be extracted as feature quantities. As another example, step S110 may include receiving a detected edge of an image (preprocessed data) or unprocessed image. Step S110 may include additional or alternative (potentially preprocessing) steps of performing processing of evidence data. Evidence data processing preferably includes any type of data processing that transforms the data into a format suitable for processing by the LBD. Some examples of evidence data processing on images may include edge detection, resolution reduction, contrast enhancement. Some examples of evidence data processing for speech may include pitch detection, frequency analysis, or generation of mel frequency cepstrum coefficients. Evidence data can be received from any suitable source. In some cases, the evidence data includes output data from the LBD. An example of such evidence data may include a dataset containing information about the characters contained within the image (eg, the output of the neural network of FIG. 7).

[0050] When evidence data is received (and processed as needed), the evidence data is preferably sent to, entered, or directed to the input of the LBD. In the case of RCN, the evidence data is preferably directed to the lowest subnetwork level child feature node. Additional or alternative, evidence data may be directed to any LBD node or connection (eg, data previously processed through levels 1 and 2 of the LBD may be inserted directly into level 3). ..

Step S120, which sets the conversion configuration, functions to provide the LBD with instructions on how the forward and reverse transformations are performed. Step S120 preferably determines when and where the forward and reverse transformations of the LBD (ie, steps S130 and S140, respectively) are performed and output is produced. Step S120 can include, for example, a step of setting a static configuration given an input to the first layer (L1) of a five-layer system, step S120 telling the LBD to perform the transformation according to: Can be instructed.
Input → L ₁ → L ₂ → L ₃ → L _{2 →} L ₃ → L ₄ → L ₃ → L ₄ → L ₅ → Output

[0052] This static configuration preferably applies both forward and reverse transformations, as seen in the above examples. The static configuration can be fully defined as to how the input-to-output conversion is applied. Alternatively, the static configuration can be a transformation pattern that covers a subset of the layers that can be invoked or executed in response to some trigger, as in the case of the dynamic transformation configuration described below. The reverse transformation can be shallow (for example, only one or a few abstraction layers travel in the opposite direction), but the reverse transformation can be deep (even back to the starting layer, many abstractions). Back layer). Further, patterns that advance the net progression of one layer include, for example, three times in the reverse direction, two times in the forward direction, one time in the reverse direction, two times in the forward direction, one time in the reverse direction, and forward. A mixed pattern can be defined by a series of transformations such as twice. Alternatively, any suitable conversion pattern can be used in any suitable combination.

[0053] Additional or alternative, step S120 may include setting a dynamic transformation configuration. This is, for example, performing recursion based on a probability threshold (for example, performing recursion until the maximum probability of the distribution exceeds 0.8 or the maximum number of recursion cycles is reached). As another example, step S120 may perform recursion based on convergence. This is, for example, performing recursion until the difference between high-level outputs falls below a certain threshold over the recursion cycle, or until the maximum number of recursion cycles is reached.

Although the examples presented herein describe recursion between layers, step S120 may additionally or optionally include the step of setting recursion between nodes or between other LBD components. .. If the LBD is RCN, step S120 may additionally or alternatively include instructions for propagation for lateral constraints.

Step S130, which performs the forward transformation, functions to infer an interpretation or classification from the evidence data. Inference can include other uses, including pattern detection, classification, prediction, system control, decision making, and inference of information from data. Forward transformations are preferably performed layer by layer within the LBD (ie, simultaneously across layers), but additionally or alternatively, subnetworks, nodes, or other suitable methods. May be good. In the case of the example shown in FIG. 7, each subnet of layer 1 receives the image data of the related section, calculates the likelihood distribution over the character and the position space, and outputs the likelihood distribution to perform forward conversion. To execute. Similarly, the subnet O (ie, layer 2) receives the likelihood distribution as evidence and calculates the likelihood distribution of the string across the string space (the string space, i.e. the space of the possible strings). Perform forward conversion by outputting the string with the maximum likelihood. The forward transformation may be performed by a subnetwork as shown in FIG. 7, or by a node, or by another suitable LBD component or substructure.

Step S130 preferably performs a step of receiving evidence at the input (eg, node, subnet, or layer) of the LBD unit and performing a mathematical transformation of the input evidence (eg, probability distribution or). Includes a step (calculating the ratio) and a step of outputting the transformed evidence to the output of the LBD unit. The mathematical transformation performed in step S130 preferably calculates the posterior probability distribution of the LBD unit based on the updated likelihood data received, but is optional, additionally or optionally, as part of the transformation. You may calculate the appropriate mathematical function of. Step S130 preferably utilizes stochastic propagation techniques to communicate information, but other stochastic inference approaches can be implemented in alternatives. Belief propagation involves exchanging messages between nodes and performing calculations within the nodes under different assumptions.

[0057] Step S130 preferably includes performing forward conversion based on the conversion configuration set in step S120.

An exemplary network is as shown in FIG. The prior probabilities P ₁ (S ₁ ), P ₁ (R ₁ ), P ₁ (Q ₁ ) and the likelihood relations P (e | S ₁ ) and P (S ₁ ) are encoded in the network. | R ₁ ), P (R ₁ | Q ₁ ). When evidence e is introduced into the system, it is first entered into S1. Considering the general form of evidence e, it can be described as follows.

[0059] While this sum is valid for a discrete probability distribution over e, one of ordinary skill in the art will recognize that it can be generalized to a continuous probability distribution. A simple example of e taking a specific value is as follows.

After calculating the posterior probability of S1, (P2 (S1)), this posterior probability is transmitted from S1 to R1 where it is used to update the posterior probability of R1 (essentially, of S1). The posterior ratio of S1 to posterior is a function used to correct or weight the likelihood p (S1 | R1)). The following is the derivation of the relationship between the posterior probabilities of R1 and the posterior probabilities of S1.

From this derivation, it is clear that posterior probabilities can be calculated in R1 by considering only the ratio of posterior probabilities to prior probabilities in S1. Similarly, it is shown that this relationship holds for Q1 (only the transmission of the ratio before and after R1 is required, not before / after S1 or evidence).

[0062] Q1 is as follows.

[0063] The network of this embodiment preferably says that calculations at any layer or node of the network depend directly only on the values output by adjacent LBD units (eg, subnetworks, nodes, layers). Shows a particular type of forward transfer to emphasize the facts. The forward conversion of S130 preferably outputs a function that depends directly on only the unit on which the conversion is performed, but may additionally or alternatively output any suitable function. Direct dependence preferably reduces the recalculation of the LBD by the unit and allows for easier reuse of the unit structure (eg, connecting many identical subnetworks together to form the LBD).

Although the previous embodiment includes explicit passing of a likelihood update message (ratio before and after, or mathematically related term), step S130 correlates directly with the message to which the likelihood or related concept is passed. Forward conversions can also be performed in non-networks (eg, binary output based on threshold functions). As shown in FIG. 10, a neural network with binary nodes calculates whether a given 4-bit number is a prime number. As shown in FIG. 11, the neural network receives an input of 0xb1011 (decimal 11). The first layer of nodes then computes the response and propagates it along the network (ie, performing forward transformations at layer 1). The second layer of the node then receives the input, computes the response, followed by the third layer of the node. Finally, the system outputs a 1, which indicates that 11 is actually a prime number.

[0065] This embodiment is somewhat limited by the output function of the node. As explained, it is possible to output only whether a certain number is a prime number. In many cases, it is useful to know the posterior probabilities (eg, the probabilities that the numbers are prime, given some evidence) instead. In the system of this embodiment, for example, it is not clear how to calculate the probability that a 4-bit binary number with the smallest digit bit of 1 is a prime number. One way to calculate this probability is to perform a number of forward paths on the network over time. To calculate the probability that a 4-bit binary number with a minimum digit bit of 1 is a prime number, the system simply has a random binary variable with a minimum digit bit of "1" and p (X = 1) = 0.5. Can be provided as input. The probability distribution can be estimated by the output of the system after some forward paths.

In the example shown in FIG. 12, method 100 performs a forward transformation of the RCN to infer the pattern of the image. The message in this embodiment represents the likelihood of evidence, assuming that the node corresponding to the source of the message is on. For example, node CF2 has a higher likelihood compared to node CF1 because the representation of node CF2 is better 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 belief-propagating 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.

[0067] Propagation is preferably initiated when a data feature input is received at the final child feature node of the network. This last child feature node is the lowest child feature node in the hierarchy. The data is preferably processed, transformed, or split 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 a feature is used to activate or not activate 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 of likelihood for which features are apparent in the data. Belief propagation continues to propagate this input up the network hierarchy. Within the subnetwork, node activation propagation involves sending the likelihood score to the PSCF node to which the child feature node is connected, and the posterior distribution component and the likelihood score of the connected PSCF node at the pool node of the subnetwork. It includes a step of generating a likelihood score from, and a step of generating a likelihood score from the posterior distribution component in the parent feature node of the subnetwork and the likelihood score of the pool node connected to the parent feature node. This belief propagation preferably continues to higher subnetworks until the network propagation progresses to the end or some threshold is met (these constraints are preferably set in step S120).

[0068] When used in an RCN, step S130 involves enforcing a selection constraint on at least the second node, which allows the invariant relationship between the pool and the subnetworks to be defined and used during inference. Works like. When a node is activated, it is preferred that the constraints imposed on it be enforced on other nodes connected through the constraint node. The external constraint node is preferably between at least two PSCF nodes, or may 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.

[0069] Step S130 preferably outputs the transformed evidence at the output of the LBD unit, which output preferably functions to process or assimilate the activated nodes of the network into the inference results. Preferably, the parent feature node is used as an indicator of the pattern. When constructing a network, it is preferable that different layers detect patterns on different particle size scales. At lower levels, this can include the detection of specific pixel patterns such as corners, lines and points. At a high level, this is 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 be present 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 may instead 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 that there are three tables, men, women, and various other coffee shop fixtures in the image.

[0070] Step S140, which performs the reverse transformation, functions to predict evidence data from the knowledge of the LBD. Additional or alternative, step S140 may include predicting evidence data based on the constraints imposed during the reverse transformation. The reverse transformation is called generation, and the special case where no external evidence is provided to the LBD is called imagination. The generation can include still images, video graphics, audio media, text content, selection behavior or response, or any suitable medium synthesized based on high level input.

[0071] During step S140 of performing the reverse transformation, the node acts on that information and propagates node selection / processing along the LBD hierarchy until the output is obtained from the output of the lowest layer subnetworks. It is preferable to let it. More specifically, the top-level subnetworks generate samples at the same time. The output sample of the top-tier subnetworks determines which lower-tier subnetworks to activate. Next, samples are generated simultaneously from the lower layer subnetworks. This output determines the active subnetwork below. This pattern continues through each layer of the LBD from the bottom layer of the subnet until a final sample is generated. In generation, the output is preferably a simulated output. For example, if LBD 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.

[0072] Similar to step S130, the reverse transformation S140 preferably occurs layer by layer within the LBD (ie, simultaneously across layers), but additionally or alternatively, per subnet, per node, or otherwise. It may occur in an appropriate manner. In the case of the embodiment shown in FIG. 13, the input character string is supplied to the output of the LBD. Given an input string, the subnetwork O performs a reverse transformation based on the input string through the calculation of probability distributions for the various features of layer 2. Similarly, layer 1 performs reverse transformation by converting intermediate data into predicted image data. This predicted image data can be thought of as the output of a set of random variables with a probability distribution defined by the LBD.

[0073] In step S140, a step of receiving a constraint condition at the output of the LBD unit, a step of executing a mathematical transformation on the information stored in the LBD to which the constraint condition is given, and a step of executing the generated data to the LBD unit. Includes steps to output at the input of. The mathematical transformation performed in step S140 preferably calculates the updated likelihood distribution for the LBD unit based on constraints, but additionally or alternatively, any suitable mathematical function as part of the transformation. May be calculated.

[0074] Step S140 preferably utilizes a probability propagation technique for communicating information, but other stochastic reasoning techniques may be implemented instead. Belief propagation involves exchanging messages between nodes and performing calculations within the nodes under different assumptions.

[0075] It is preferable that step S140 includes a step of executing reverse conversion based on the conversion configuration set in step S120.

An exemplary network is as shown in FIG. The prior probabilities P ₁ (S ₁ ), P ₁ (R ₁ ), P ₁ (Q ₁ ) and the likelihood relations P (e | S ₁ ) and P (S ₁ ) are encoded in the network. | R ₁ ), P (R ₁ | Q ₁ ). If the constraint Q1 = q is introduced into the system, it can be plugged directly into the known likelihood and used as the updated probability of R1.
P ₂ (R ₁ ) = P (R ₁ | Q ₁ = q)

Similarly, S1:

この確率分布は、与えられた制約Ｑ１＝ｑについてＬＢＤで予測された証拠の分布を表している。Ｑ１にわたって合計することによって、層が大きなＬＢＤのどこに存在するかに拘わらず、各層について可能性のあるすべての出力をその層への入力関数として計算することができる。ＲＣＮの場合、あらゆる可能なオブジェクト（潜在的にＱ１＝｛ｑ１、ｑ２、…｝によって表される）は、各層の計算についてグラフへと拡張され、これによりプール選択の問題を因子グラフとして公式化することが可能となる。因子グラフの因子のパラメータは入力に依存するが、ＲＣＮのより大きな構造は依存しない。この因子グラフを事前計算することにより、任意の所望のオブジェクトについて、更新された最大積の順序付けと割り当てを格納することが可能になり、それによって高速なオブジェクト認識が可能になる。これは静的逆方向変換と呼ばれ、ステップＳ１４０に含まれてもよい。 [0078] Further, a probability distribution describing e may be generated.

This probability distribution represents the distribution of evidence predicted by the LBD for the given constraint Q1 = q. By summing over Q1, all possible outputs for each layer can be calculated as input functions to that layer, regardless of where the layers are located in the large LBD. In the case of RCN, every possible object (potentially represented by Q1 = {q1, q2, ...}) Is extended to a graph for the computation of each layer, thereby formulating the pool selection problem as a factor graph. It becomes possible. The parameters of the factors in the factor graph depend on the input, but not on the larger structure of the RCN. Pre-computing this factor graph makes it possible to store updated maximum product ordering and allocation for any desired object, which allows for faster object recognition. This is called static reverse transformation and may be included in step S140.

[0079] Step S140 may additionally or optionally include a step of performing a dynamic reverse transformation. Unlike static reverse transformation, where the output is a probability distribution based on predicted activation with some constraints, dynamic reverse transform directly activates the features of the constrained LBD and is an example output (or). , In the case of repetition, a set of output examples) is generated. This preferably allows the detection of new objects and / or generalization of the behavior of parts of the object.

[0080] In this network embodiment, it is preferred that calculations at any layer or node of the network directly depend only on the values output by adjacent LBD units (eg, subnetworks, nodes, layers). To emphasize the facts, a particular type of inverse transformation is shown. The reverse conversion of S140 preferably outputs a function that directly depends only on the unit on which the conversion is performed, but may additionally or alternatively output any suitable function. Direct dependence preferably reduces recalculation by the LBD unit and allows easier reuse of the unit structure (eg, using many identical subnetworks connected together to form the LBD). To do).

[0081] The previous embodiment includes explicit passing of a probability update message (likelihood calculation or mathematically related term), whereas step S140 includes a message (eg,) in which a probability or related concept is passed. Inverse transformations may be performed in networks that do not directly correlate with (binary output based on the threshold function).

In the example shown in FIG. 15, Method 100 performs a reverse transformation on the RCN to generate a pattern. 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. First, the pattern parent feature input is received. The parent feature is preferably another input on which a high level feature, classification, or pattern is generated. The input is preferably provided to the top subnetwork of the network. Next, propagation over the network proceeds. Processing continues so that the top layer subnetworks are processed, then the next layer of the subnetwork is processed, and each layer of the network is processed incrementally (ie, sequentially or continuously). In some cases, the external constraint defines the relationship between the two subnetworks, one subnetwork is processed first, then the other subnetworks is processed with the external constraint in mind. This order may be predefined or configured. Alternatively, the process can be a race condition between different subnetworks and the first subnet network to complete the process of determining the enforcement of the constraint. Alternatively, they may be processed or managed simultaneously in any suitable manner. Similarly, there may be an order of processing between nodes in the subnetwork. The pools in the subnetwork are preferably ordered in the same way. In some cases, the lateral constraint defines the relationship between the PSCF nodes of the two pools, one pool is processed first, then the other pool is processed with the lateral constraint in mind. The pool. This order may be predefined or set. Alternatively, the process can be a race condition between the different pools and the first pool to complete the process and determine the enforcement of constraints on the other pools. Alternatively, they may be processed or managed simultaneously in any suitable manner. Within each subnet, node selection begins with the parent feature node, then the pool node is activated, and the PSCF node is selected. The selection of PSCF nodes can be at least partially affected or determined by the enforced selection constraints of the constraint node. Choosing a pool node that is consistent with the function of the parent feature node works to properly activate the pool in the subnetwork. 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. This choice function is preferably in an AND relationship such that each connected pool node is activated, but any suitable choice function can be used instead.

[0083] Selecting at least the first PSCF node corresponding to a child feature of a subnetwork functions to select a PSCF node within the set of pool elements of the pool node. Selection is made for each selected pool node. The order in which pool nodes in a subnetwork are evaluated may be random, continuous, and non-simultaneous. Alternatively, the 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 selection function may be used. The PSCF node is preferably connected to or otherwise associated with at least one child feature node in a direct relationship-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 higher-level connection to the child feature node.

Enforcing selection constraints serves to define invariant relationships between pools and subnetworks. Constraints are preferably created to define the logic between feature combinations and patterns. In a typical example, subnetworks splice image components together to form an image of a car, and if one pool selects the body of the car, then the wheels of the car are selected in the other pool. Limits can be imposed, so the wheels and car body are consistent. Selection constraints can be defined via connections between at least two PSCF nodes via constraint nodes. The constraint node has any suitable number of connected PSCF nodes and may perform any suitable selection function. In some cases, selection constraints can be defined via a connection between two pool nodes or any suitable type of node. Similarly, constraint nodes can be between any two or more types of nodes, such as between PSCF nodes and pool nodes. Execution of constrained nodes preferably has some form of orientation at run time, for example, such that the selection of the first node has the effect of the selection on the second node. This direction can also go in any direction between the two types of nodes. The PSCF node may result in the constraint node affecting the pool node, and the pool node may 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 this constraint node is an AND operation. The selection constraint is preferably enforced in response to the selection of at least the first PSCF node having the connected constraint node. As mentioned above, the nodes are preferably evaluated or propagated in some order. The selection constraint preferably does not apply to PSCF nodes that have already been selected, but instead applies to selection by pool nodes. In some scenarios, one pool node has a set of possible PSCF nodes, which are reduced to one node after the selection constraint is enforced and sent to the pool element through the constraint node. .. In other scenarios, pool nodes may reduce the number of possible PSCF nodes or change the probabilistic weighting for selection. One constraint node is shown as a connection between two PSCF nodes, but these constraints can also be implemented operably via a pool element and / or message passing mechanism between subnetworks as an alternative. it can. The message preferably modifies the behavior of the choice function to effectively enforce the constraint node as described herein. Constraint nodes can be lateral constraints, external constraints, time constraints, and / or any suitable type of constraint. Lateral constraints are preferably enforced between two different pools. External constraints are preferably enforced 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 for different time instances. Time constraints can define invariant patterns over different time frames. Temporal selection constraints identify features that may or may not occur within a set of features. Compile the final child features of the network into a generated output function and assemble the features into the product, representation, or analysis, simulation, or any suitable output that is generated. The final child feature is preferably the child feature node at the bottom of the hierarchical network. The child feature node preferably represents a binomial variable that represents the presence of a particular data feature. You may have a database or mapping that maps child feature nodes to specific data features. The step of compiling the final child feature preferably includes the step of mapping the selected child feature node to the data feature, which is compiled into the next generated output. Activated child feature nodes are preferably components that, when combined, form a replica of the media. For example, if the network is trained or created for image generation, the output is preferably a substantially complete simulated image. If the network is trained with audio features, the final child features can be combined to output an audio file or signal. If multiple network evaluations are used for the time signal, the final child features of the multiple networks can be compiled into the final output.

The method 100 may include performing steps S130 and S140 at any time and at any position on the LBD, preferably according to the conversion configuration of step S120. For example, the LBD can perform a series of partial forward and reverse transformations, as shown in FIG. As another example, the LBD may acquire an image input for half of the image. This semi-image input is input to the child feature. The LBD is then prompted to create a possibility for the other half.

[0086] Step S150 includes a step of outputting the converted evidence data, and functions to output the data interpretation generated by the LBD. The step of outputting the converted evidence data preferably includes a step of post-processing the output, but may additionally or optionally include a step of outputting the unprocessed output data. For example, the output data may include a set of classification tags for the image, post-processed from the probability distribution across the classification tags. As another example, S150 may include the step of outputting a natural language description of the object in the photo.

[0087] The methods of the preferred embodiment and variations thereof can be implemented and / or implemented at least partially as a configured device that accepts a computer-readable medium for storing computer-readable instructions. The instructions are preferably executed by computer-executable components integrated with the 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 provide alternative or additional instructions. ..

[0088] As will be appreciated by those skilled in the art from the above detailed description as well as the 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.

Claims (20)

A method for generating data interpretations in a two-way layer-based network, where computer-executable components
-Steps to receive a set of evidence data and
· Steps to set up a transformation configuration that directs the messaging of evidence and transformed data between layers of the network,
A step of performing a first forward conversion of a set of evidence data into a first set of converted data in the first layer of the network.
A step of performing a second forward conversion from the first set of conversion data to the second set of conversion data in the second layer of the network.
In the second layer of the network, a step of performing a first reverse conversion of the second set of converted data into a third set of converted data,
A step of performing a third forward conversion of the third set of conversion data into a fourth set of conversion data in the second layer of the network.
-The step of outputting the fourth set of conversion data
A method characterized by performing . The method according to claim 1, wherein the bidirectional layer-based network is a recursive cortical network. In the method of claim 1, the step of executing the second forward conversion is
○ A step of receiving the first set of conversion data at one or more inputs of the second layer, and
○ A step of calculating a first posterior probability distribution based on the first set of conversion data and a set of prior probabilities and likelihood relationships encoded in the second layer.
○ The output of the second layer includes a step of generating the second set of conversion data from the first posterior probability distribution.
The step of performing the first reverse transformation is
○ In the output of the second layer, the step of receiving the second set of conversion data and
○ A step of calculating an updated set of likelihood based on the second set of transformation data and the set of prior probabilities and likelihood relationships encoded in the second layer.
A method comprising: in one or more inputs of the second layer, a step of generating the third set of conversion data from the updated set of likelihoods. In the method of claim 3, the step of executing the third forward conversion is
○ A step of receiving the third set of conversion data at one or more inputs of the second layer, and
○ Based on the first set of transformation data, the updated set of likelihood of the third set of transformation data, and the set of prior probabilities and likelihood relationships encoded in the second layer. Then, the step of calculating the second posterior probability distribution,
○ A method characterized in that the output of the second layer includes a step of generating a fourth set of conversion data from the second posterior probability distribution. The method according to claim 4, wherein the bidirectional layer-based network is a recursive cortical network. The method of claim 5, wherein the recursive cortical network is performed by a distributed processing element . In the method of claim 5, the recursive cortical network is
· Includes recursively designed networks of sub-networks organized into multiple layers
-The sub-network includes at least a parent feature node, a pool node, a parent-specific child feature node, and a child feature node.
Parent feature node of the at least one sub-network is constituted by the at least one sub-network operable selection function on at least two pools nodes connected to the parent feature node,
• The pool node of at least one subnetwork is composed of a choice function that can operate on at least two parent-specific child feature (PSCF) nodes connected to the pool node of at least one subnet .
The PSCF node of at least one of the subnetworks is configured to activate the connected child feature node.
- child feature node can be connected to at least a parent feature node of the second sub-network of the lower layer,
A method comprising a constraint node having at least two connections from at least two PSCF nodes and having a choice function for enhancing selection by said pool node. The method according to claim 5, wherein the set of evidence data includes image data. The method according to claim 8, wherein the image data includes image data processed by an edge detection filter. The method according to claim 9, wherein the image data is captured by a camera, and the step of outputting the fourth set of conversion data includes a step of outputting interpretation data of the image. A method of claim 4, further comprising performing additional forward and reverse transformations based on the static transformation configuration. The method of claim 4 further comprises performing additional forward and reverse transformations based on the dynamic transformation configuration, wherein the dynamic transformation configuration is based on a layer output likelihood threshold. A method characterized by directing messaging. The method of claim 4 further comprises performing additional forward and reverse transformations based on the dynamic transformation configuration, wherein the dynamic transformation configuration directs messaging based on a recursive level threshold. A method characterized by doing. A method of claim 4, further comprising performing a forward transformation based on a lateral constraint encoded in the network. A method according to claim 14, wherein the step of executing the forward transformation based on the lateral constraint includes a step of performing a selection constraint between at least two nodes of the network. The method according to claim 4, wherein the step of outputting the fourth set of conversion data includes a step of outputting a set of data classifiers. A method of generating data interpretations in a recursive cortical network, a computer-executable component of
A step of receiving a set of evidence data at a child feature node of the first layer of the recursive cortex network, the recursive cortex network consisting of a set of layers, each layer having at least two child feature nodes and one. Steps with one parent node,
A step of setting a conversion configuration that directs the messaging of evidence data and converted data between layers of the network, and
A step of performing a series of transformations, including at least one forward transformation and at least one reverse transformation, on the evidence data according to the transformation configuration.
・ Steps to output the converted evidence data
A method characterized by performing . A method of claim 17, wherein the conversion configuration is a static conversion configuration that explicitly enumerates the order of forward and reverse transformations performed by the method. The method according to claim 17, wherein the conversion configuration is a dynamic conversion configuration that sets the order of forward conversion and reverse conversion based on the probability threshold value of the layer output. A method according to claim 17, wherein the conversion configuration is a dynamic conversion configuration that sets the order of forward conversion and reverse conversion based on a threshold value of a recursive level.

Images