WO2023210816A1 - Information processing device, information processing method, and program - Google Patents

Abstract

The present invention addresses the problem of improving convenience when recognizing spatio-temporal context in information processing.　A single layer neural network 12 comprises a feedforward circuit 121 having 120 connection weights WS and a feedback recursive circuit 122 having 1 to 4 recursive connection weights WK, wherein when the storage of spatio-temporal context is to be learned, learning using a spatio-temporal learning rule is applied to the connection weights WS of the 120 or more circuits of the feedforward circuit 121, while learning using a Hebbian learning rule is applied to the recursive connection weights of the 4 circuits of the feedback recursive circuit 122. This solves the above-described problem.

Description

Information processing device, information processing method, and program

The present invention relates to an information processing device, an information processing method, and a program.

Conventionally, there is a technology that uses a multilayer neural network to generate (learn) a model that performs image recognition (for example, see Patent Document 1).

JP2015-095215A

Prior art such as the above-mentioned patent document 1 and AI technology using deep learning are capable of memorizing spatiotemporal context (recognition ) was being carried out. However, in such conventional spatio-temporal context recognition, sufficient needs such as energy cost and separation (accuracy) of similar spatio-temporal contexts are not met, and improvements in convenience have been desired.

The present invention has been made in view of these circumstances, and aims to improve convenience in recognizing spatio-temporal context in information processing.

To achieve the above object, an information processing device according to one embodiment of the present invention includes:
A one-layer neural network comprising a feedforward circuit having one or more connection weights and a feedback circuit having one or more recursive connection weights,
When performing memory learning for spatiotemporal context,
Performing learning applying a spatio-temporal learning rule to the one or more connection weights of the feedforward circuit,
performing learning applying Hebb's learning law to the one or more recursive connection weights of the feedback circuit;
The one-layer neural network is provided.

An information processing method and program according to one embodiment of the present invention are methods and programs corresponding to an information processing apparatus according to one embodiment of the present invention.

According to the present invention, it is possible to improve the convenience in recognizing spatiotemporal context in information processing.

1 is a diagram illustrating an overview of a configuration example of an embodiment of an information processing device of the present invention. 2 is a diagram illustrating an example of input data and output data in the information processing apparatus of FIG. 1. FIG. FIG. 2 is a diagram showing an overview of the characteristics of the feedforward circuit and feedback circuit shown in FIG. 1. FIG. FIG. 2 is a diagram showing an overview of the characteristics of the feedforward circuit and feedback circuit shown in FIG. 1. FIG. 2 is a diagram showing a configuration focusing on a feedforward circuit out of the configuration of FIG. 1. FIG. 5 is a schematic diagram of a neuron showing characteristics of a spatio-temporal learning rule in the feedforward circuit of FIG. 4. FIG. 2 is a diagram showing a configuration focusing on a feedback recursive circuit out of the configuration of FIG. 1. FIG. FIG. 9 is a diagram illustrating characteristics of the Hebbian learning rule in the feedback recursive circuit of FIG. 8; FIG. 9 is a diagram illustrating characteristics of the Hebbian learning rule in the feedback recursive circuit of FIG. 8; 10 is a diagram illustrating a physiological experiment in which the Hebbian learning law and the spatio-temporal learning law of FIG. 9 coexist. FIG. 11 is a diagram showing changes in long-term enhancement of synaptic load in the physiological experiment of FIG. 10. FIG. FIG. 2 is a diagram showing an example of a computer simulation using the one-layer structure feedforward circuit and feedback recursive circuit of FIG. 1 and its results. 13 is a diagram showing an example of the relationship between learning speed parameters and learning results in the computer simulation of FIG. 12. FIG.

Embodiments of the present invention will be described below with reference to the drawings.
FIG. 1 is a diagram showing an overview of a configuration example of an embodiment of an information processing apparatus of the present invention.
The information processing device 1 includes an input unit 11, a neural network 12 to which one aspect of the present invention is applied (hereinafter referred to as “this neural network 12”), and an output unit 13.

Although details will be described later, this neural network 12 is a one-layer neural network composed of a feedforward circuit 121 having a connection weight K of 1 or more and a feedback recursion circuit 122 having a recursion weight WS of 1 or more. It is.
By using the present neural network 12, it is possible to efficiently memorize spatio-temporal contexts to be described later.

The input unit 12 inputs the input data ID to the neural network 12.
The output unit 13 outputs output data OD from the neural network 12 when the input data ID is input.

Here, the input data ID is the following data.
First, at a predetermined timing (instantaneous), one unit of data DA (hereinafter referred to as "unit data DA") consisting of a plurality of bits is simultaneously input from the input unit 12 to the present neural network 12.
Hereinafter, it is assumed that data in a 10×12 matrix consisting of 10 rows is employed as the unit data DA, with 12 bits of data representing one row.

Then, a plurality of unit data A having different contents are sequentially input from the input unit 12 to the neural network 12 in time.
Here, K pieces of different unit data A are set as one set, each of the K pieces of unit data A is arranged in a predetermined order, and the unit data A is arranged one by one at a predetermined time interval in the order of arrangement. Each signal is input to the neural network 12 from the input unit 12.
Note that hereinafter, a pattern in which K different unit data A are arranged in the order in which they are inputted in time will be referred to as a "temporal pattern." Note that for convenience of explanation below, K=5.

L patterns are prepared as time patterns.
That is, each of the L time patterns is sequentially input from the input section 12 to the present neural network 12.
Here, as shown in FIG. 1, the time direction is set as the left-right direction, and the time pattern that is inputted first in time is placed at the top, and then the time pattern that is inputted to the present neural network 12 is arranged in the order from top to bottom. If you place the time patterns in , and place the time pattern that is input last in terms of time at the bottom, a K×L matrix consisting of L rows is created, with K pieces of unit data A (time pattern) as one row. data is generated. The data in this L×K matrix is the input data ID.
If we fix the time (temporal order) and focus on the arrangement of L unit data A in the vertical direction, the pattern of L unit data A in the vertical direction will be determined at each time (temporal order). You can see that it's different. Note that, hereinafter, such an up-down direction will be temporarily referred to as a "spatial direction", and a pattern of L pieces of unit data A in this spatial direction will be referred to as a "spatial pattern". Note that for convenience of explanation below, L=24.

In this way, the input data ID (L×K unit data DA) is understood to be composed of L temporal patterns and K spatial patterns. It can be seen that a plurality of patterns (patterns in which the unit data A, which is each element of the matrix, is different) can be prepared for such input data ID (L×K unit data DA). Therefore, the pattern of the input data ID (L×K unit data DA) is called a "spatiotemporal pattern."

In summary, as shown in FIG. 2, the present neural network 12 outputs output data OD based on input data ID.
Hereinafter, an example of the above-mentioned input data ID and output data OD will be explained using FIG. 2.
FIG. 2 is a diagram illustrating an example of input data and output data in the information processing apparatus of FIG. 1.

Here, the input data ID is the following data.
The input data ID is composed of L×K unit data DA, and has L temporal patterns and K spatial patterns.
The unit data DA (data A3 in the example of FIG. 2) is data consisting of 120 bits. This 120-bit data is simultaneously input from the input unit 11 to the neural network 12 at a predetermined timing.

Then, a plurality of unit data A having different contents are sequentially input from the input unit 12 to the neural network 12 in time. Here, each unit data has a Hamming distance H from each other. The Hamming distance H refers to the number of characters that differ between two character strings (here, 120-bit bit strings).
The neural network 12 can recognize the time-series context of input data ID consisting of five unit data DA with minute differences, such as Hamming distance H=8 bits out of 120 bits, for example.

In this way, the input data ID (L×K unit data DA) is understood to be composed of L temporal patterns and K spatial patterns. It can be seen that a plurality of patterns (patterns in which the unit data A, which is each element of the matrix, is different) can be prepared for such input data ID (L×K unit data DA). Therefore, the pattern of the input data ID (L×K unit data DA) is called a "spatiotemporal pattern."
In this way, the time pattern refers to a pattern in which K=5 unit data A, each having a different content (meaning), are arranged in the order in which they are input. For the human brain (in this embodiment, the neural network 12), this temporal pattern is understood as a so-called context.
Input data composed of L temporal patterns and K spatial patterns has a spatiotemporal pattern, so the configuration of the input data ID is appropriately called a spatiotemporal context pattern.

When one time pattern (5 unit data DA) is input to the neural network 12 from the input section 11, data DP in units of 120 bits (hereinafter referred to as unit data DP) is output.
That is, when L=24 time patterns are sequentially input in the arrangement order shown in FIG. 2, L=24 unit data DP as shown in FIG. 2 are arranged in the spatial direction, and this becomes output data OD. In this way, the output data OD has a spatial pattern corresponding to the spatiotemporal pattern of the input data ID.

Here, the feedforward circuit 121 and feedback recursive circuit 122 included in the present neural network 12 will be briefly explained.

The feedforward circuit 121 is a circuit that applies a spatio-temporal learning rule.
Although details will be described later, the spatio-temporal learning rule has the property shown in the following equation (1).

...(1)

The feedback recursive circuit 122 is a circuit that applies Hebb's learning rule.
Although details will be described later, the Hebbian learning rule has the property shown in the following equation (2).

...(2)

The feedforward circuit 121 and the feedback recursive circuit 122 are coupled at a coordination ratio α, as shown in equation (3) below.

...(3)

As a result, the spatio-temporal context pattern of the input data ID is output as a difference in the spatial pattern of the output data OD depending on the spatio-temporal context pattern. In other words, the present neural network 12 is a neural network that recognizes spatio-temporal context patterns and can output different output data OD depending on the spatio-temporal context patterns of input data ID.

Focusing on an instant (predetermined time), each of the 120 bits constituting the unit data A is input from the input unit 11 to the neural network 12 at the same time. That is, when simulating the human brain, it can be understood that the input section 11 corresponds to a region made up of 120 input cells.
The connection weight WS and the recursive connection weight WK in the neural network 12 can be understood to correspond to a feedforward synaptic load and a feedback synaptic load, respectively, when imitating the human brain.
In addition, it can be understood that the output unit 13 corresponds to a part made up of 120 output cells when compared to a human brain.

Next, the feedforward circuit 121 having one or more coupling loads WS and the feedback recursive circuit 122 having one or more recursive coupling loads WK will be explained using FIGS. 3 to 9.

First, the characteristics of the feedforward circuit 121 and the feedback recursive circuit 122 will be briefly described.
FIG. 3 is a diagram showing an overview of the characteristics of the feedforward circuit and feedback circuit shown in FIG. 1.

The feedforward circuit 121 has an excellent pattern separation function (ability) shown in FIG. 3A.
That is, as shown in FIG. 3A, there is an overlapping region between the input A pattern and the A' pattern. This indicates that the bit strings corresponding to the A pattern and the A' pattern at the time of input are similar to a predetermined degree.
On the other hand, the A pattern and the A' pattern of the output in FIG. 3A do not overlap and are separated from each other. This indicates that the bit strings corresponding to the A pattern and the A' pattern at the time of output are not similar.
In this way, the feedforward circuit 121 applies the spatio-temporal learning rule, and even if patterns that are similar to a predetermined degree are input, dissimilar outputs are produced.

The feedback recursive circuit 122 has an excellent pattern complementation function (ability) shown in FIG. 3B.
That is, as shown in FIG. 3B, there is an overlapping region between the input A pattern and the A' pattern. This indicates that the bit strings corresponding to the A pattern and the A' pattern at the time of input are similar to a predetermined degree.
On the other hand, the A pattern and the A' pattern of the output in FIG. 3B overlap more than the predetermined degree of the input. This indicates that the bit strings corresponding to the A pattern and the A' pattern at the time of output are extremely similar.
In this way, the feedback recursive circuit 122 applies Hebb's learning law, and even if patterns that are similar to a predetermined degree are input, extremely similar outputs will be produced.

This neural network 12 realizes excellent spatiotemporal context pattern recognition by combining a feedforward circuit 121 with excellent pattern separation function and a feedback recursion circuit 122 with excellent pattern completion function in one layer. This is possible.
Hereinafter, the characteristics of the feedforward circuit 121 having an excellent pattern separation function and the feedback recursive circuit 122 having an excellent pattern complementation function will be explained in more detail.
Hereinafter, the characteristics of the feedforward circuit 121 will be explained in more detail using the results of a physiological experiment of the spatio-temporal learning rule with reference to FIGS. 4 to 7.

FIG. 4 is a diagram showing the configuration of FIG. 1, focusing on the feedforward circuit.
As explained using FIG. 1, also in FIG. 4, certain unit data IDs of input data ID are simultaneously input to the feedforward circuit 121.
The unit data DA that is input at the same time is subjected to the spatiotemporal learning rule of ΔW ^S _ij shown in the above equation (1) in the feedforward circuit 121 that is coupled with the feedforward. As will be described in detail later, the parameter ΔW ^S _ij is the amount of change in the connection weight WS of the synapse W _ij . Hereinafter, the circuits within the neural network 12 and the connections between the circuits will be appropriately referred to using the term synapse in association with the human brain.

The characteristics of the feedforward circuit 121, that is, the spatio-temporal learning rule will be explained.
FIG. 5 is a schematic diagram of a neuron showing the characteristics of the spatio-temporal learning rule in the feedforward circuit of FIG. 4.

The spatiotemporal learning law induces synaptic load changes (plasticity) depending on the synchronization rate (I _ij (t)) between input cells (synapses).
Specifically, as shown in FIG. 5, a signal having the weight of the synapse W _ij is input from the i-th input x _i at a predetermined timing shown at times t...t _-m . At this time, for example, as shown in FIG. 5, a signal of the weight of the synapse _Wkj is also input from _xk .
The spatio-temporal learning shown in the above equation (1) is performed on the signals input in this way.
Note that in equation (1), h(x) is as shown in equation (4) below.

...(4)

As a result, as shown in FIG. 5, three types of load changes occur depending on the synchronization rate at which synchronization from a plurality of synapses occurs. When the synchronization rate is high (x is θ ₁ or more), there is a positive load change, when the synchronization rate is low, there is a negative load change, and when the synchronization rate is intermediate (x is θ ₂ or less), there is a negative load change. When is intermediate, both (x is smaller than θ ₁ and larger than θ ₂ ) remain unchanged.

In this way, the spatiotemporal learning rule induces changes in synaptic weights based on the synchrony of firing between input cells, and has excellent ability to separate patterns in spatiotemporal context. Furthermore, unlike the feedback recursive circuit 122 described later, the spatiotemporal learning rule has a property that it is not directly related to the firing of the output cell.
That is, the feedforward circuit 121 senses the difference between the input spatiotemporal patterns and exhibits the pattern separation function (ability) shown in FIG. 3A.

Next, physiological experiment results regarding the spatiotemporal learning rule applied to the feedforward circuit 121 will be explained.

For details of this experiment, see Tsukada M, Aihara T, Kobayashi Y, Shimazaki H: Spatial analysis of spike-timing dependent LTP and LTD in the C. A1 area of hippocampal slices using optional imaging. Hippocampus, 15, 104-109, 2005. This will be explained with reference to the figures described in .
This figure shows the results of a physiological experiment using the hippocampal CA1 circuit (hippocampal slice experiment) of long-term enhancement of synaptic load due to synchronous firing between input cells and long-term suppression due to asynchrony.
Specifically, the paper includes slice diagrams of the hippocampal CA1 circuit and Stim. A and Stim. The positional relationship of B is illustrated.

Stim. shown in the same paper. A and Stim. An example of stimulation timing in B is illustrated.
Stim. At A, stimulate every 2s (seconds).
Also, Stim. In B, Stim. It is offset (shifted) from A by a time difference τ and stimulated every 2 s (seconds).
That is, a stimulus with τ=0 is a "synchronous" stimulus, and a stimulus other than τ=0 is an "asynchronous" stimulus.

When synchronous stimulation was applied, LTP/LTD was approximately 200%, indicating long-term potentiation (LTP).
When asynchronous stimulation was applied, LTP/LTD was approximately 100% when τ = +50ms, +10ms, -10ms, and -50ms, indicating that neither long-term potentiation (LTP) nor long-term depression (LTD) occurred. .
Furthermore, when asynchronous stimulation was applied, LTP/LTD was approximately 70% when τ = +20 ms and -20 ms, indicating long-term suppression (LTD).

As described above, experiments using the hippocampal CA1 circuit (hippocampal slice experiments) have revealed the existence of long-term potentiation (LTP) of synaptic load changes (plasticity) due to synchronous stimulation.

Above, the characteristics of the feedforward circuit 121 have been explained in more detail using the physiological experiment results of the spatio-temporal learning rule.
Hereinafter, the characteristics of the feedback recursive circuit 122 will be explained in more detail using the results of a physiological experiment in which the Hebbian learning law and the spatio-temporal learning law coexist with reference to FIGS. 8 and 9.

FIG. 8 is a diagram showing a configuration focusing on the feedback recursive circuit among the configurations of FIG. 1.
Feedback recursive circuit 122 includes the following two types of circuits. Type 1 has a recursive connection weight WS from the output cell side to the input cell side. In type 2, the firing of the output cell directly propagates back to the input side of the dendrite and has a connection weight WS.
In both cases, the Hebbian learning law is applied in that feedback from the output cell to the input side causes a load change ΔW _ij ^H that depends on the timing of output firing. .

Note that, as explained using FIG. 1, also in FIG. 8, unit data DA with input data ID are input at the same time. Moreover, as explained using FIG. 1, it works in cooperation with the feedforward circuit and the balance thereof is the emphasis ratio α. The ratio of the feedforward circuit load SK and the feedback load HK is added at a ratio of α:1−α.

The characteristics of the feedback recursive circuit 122, that is, the Hebbian learning rule, will be explained.
FIG. 9 is a diagram illustrating the characteristics of the Hebbian learning rule in the feedback recursive circuit of FIG. 8.
FIG. 9(A) shows a schematic diagram of a neuron showing the characteristics of the Hebbian learning rule in the feedback recursive circuit 122 of FIG. 8.
FIG. 9B shows an example of the characteristics of the Hebbian learning rule in the feedback recursive circuit 122 of FIG. 8, that is, long-term enhancement and long-term suppression of spike timing.

The feedback recursive circuit 122 also includes the phenomenon of spike-timing-dependent long-term enhancement and long-term suppression (spike-timing-dependent LTP and LTD, STD/LTP/LTD) due to firing of output cells.
Specifically, long-term enhancement and long-term suppression of spike timing is performed as shown in FIG. 8(B).

In this way, the Hebbian learning rule self-organizes the input information at that time depending on the firing of the output cells (later cells). Note that if the output cell does not fire, nothing will happen, that is, neither long-term enhancement nor long-term suppression will occur. The memory information processing of this circuit network is known to have the ability to represent similar items as a single pattern and a pattern complement function. In other words, a circuit network that follows Hebb's learning law will recall the entire original information even if some missing information is input. For details on the fact that a circuit network that follows the Hebbian learning rule has a pattern completion function, see Amari, S. : Learning patterns and pattern sequences by self-organizing nets of threshold elements Insti Tute of Electrical and Electronics Engineers Transactions C-21 (1972) 1197-1206. , and Hopfield, J. J. : Neural networks and physical systems with emergent collective computational properties. Proceedings of National Academy of Science (USA) 79 (1982) 2554-2558. Please refer to

In this way, the Hebbian learning rule induces synaptic weight changes based on the timing of firing spikes between input cells (previous cells) and output cells (later cells), and has excellent pattern completion ability. Furthermore, unlike the Hebbian learning law, the spatiotemporal learning law has the ability to separate similar patterns without being directly related to the firing of output cells.
That is, the feedback recursive circuit 122 using the Hebbian learning rule exhibits the common pattern convergence function and pattern complementation function (ability) shown in FIG. 3B.

Next, the results of a physiological experiment in which the Hebbian learning law and the spatio-temporal learning law that are applied to the feedback recursive circuit 122 coexist will be described using FIG. 10.
FIG. 10 is a diagram illustrating a physiological experiment in which the Hebbian learning law and the spatio-temporal learning law of FIG. 9 coexist.
FIG. 11 is a diagram showing changes in long-term enhancement of synaptic load in the physiological experiment of FIG. 10.

Figures 10 and 11 show the results of a physiological experiment in which long-term enhancement of synaptic load based on the synchrony of input stimuli according to the spatiotemporal learning law and long-term enhancement due to the firing of output cells according to the Hebbian learning law coexist. .
Specifically, as shown in FIG. 10, stimulation A and stimulation B are synchronously applied to two points 200 um apart on the axon of a neuron.
At this time, physiological experiments were conducted to measure changes in long-term potentiation of synaptic load in two experiments in which the presence or absence of blocking of back-propagation spikes in the firing of output cells was varied.

The graph shown in FIG. 11 shows the mean value (Mean) and standard error (S.E.M.) of the measurement results on a plane where the vertical axis is the long-term enhancement of synaptic load (%) and the horizontal axis is the elapsed time (minutes). , Standard Error of the Mean) is indicated by each marker.
The black markers shown in FIG. 11 are the results of measurements performed by blocking back propagation in the firing of output cells.
The white markers shown in FIG. 11 are the results of measurements in which back-propagation in the firing of output cells was not blocked.

As shown in FIG. 11, according to the results of measurements performed by blocking back propagation in the firing of output cells, the long-term enhancement of synaptic load after stimulation is approximately 120%.
That is, as a result of the measurement performed by blocking the back propagation in the firing of the output cell shown in FIG. 10, it can be seen that a long-term enhancement of the synaptic load based on the synchrony of the input stimulus occurs.

Furthermore, as shown in FIG. 11, according to the results of measurements in which back propagation in the firing of output cells was not blocked, the long-term enhancement of synaptic load after stimulation was approximately 150%. This long-term enhancement of the synaptic load is larger than the measurement result in which back-propagation in the firing of the output cell shown in FIG. 10 was blocked.
This is due to the measurement results shown in Figure 10 in which back-propagation in the firing of output cells was not blocked, and in addition to the long-term enhancement of synaptic load based on the synchrony of the input stimulus, long-term enhancement due to the firing of output cells occurs. It shows.
That is, as a result of the measurement performed by blocking the back propagation of the firing of the output cell in Figure 10, the long-term enhancement of the synaptic load based on the synchrony of the input stimulus coexists with the long-term enhancement due to the firing of the output cell of the Hebbian learning law. It can be seen that
More specifically, it can be seen that in the hippocampal CA1 circuit network (hippocampal slice experiment), the spatiotemporal learning law in the feedforward circuit 121 and the Hebbian learning law (long-term spike timing enhancement) in the feedback recursive circuit 122 coexist.
For details of this experiment, see Tsukada M, Yamazaki, Y. , Kojima, H. : Interaction between the spatiotemporal learning rule (STLR) and Hebb type (HEBB) in single pyramidal cells in the hippoca mpal CA1 area. Cogn. Neurodyn. , 1, 157-167, 2007. It is described in.
As described above, the physiological experiment shows that the feedforward circuit 121 and the feedback recursive circuit 122 coexist in the actual hippocampal CA1 circuit network.

The characteristics of the feedback recursive circuit 122 have been described above in more detail with reference to FIGS. 8 and 9 using physiological experiment results in which the Hebbian learning law and the spatio-temporal learning law coexist.
In this way, physiological experiments have shown that the properties of the spatiotemporal learning law and the Hebbian learning law coexist in one actual cell. Therefore, the inventor cooperated (combined) the feedforward circuit 121 and the feedback recursive circuit 122 in a one-layer structure, added their respective loads with a balance of cooperation degree α, and thereby memorized (recognized) spatiotemporal context patterns. I came up with the idea of doing it.

The inventor conducted a computer simulation to verify whether a spatiotemporal context pattern can be memorized (recognized) by coordinating the feedforward circuit 121 and the feedback recursive circuit 122 in a single layer structure.
FIG. 12 is a diagram illustrating an example of a computer simulation using the one-layer structure feedforward circuit and feedback recursive circuit of FIG. 1 and its results.
As shown on the left side of FIG. 12, the 120 bits of the unit data DA explained using FIGS. 1 and 2 are used as one image of 12×10 pixels as an input vector for the computer simulation, and the For the difference (Hamming distance), 10 bits were prepared.
As shown in the center of FIG. 12, 24 spatio-temporal context patterns are prepared in which the arrangement of a plurality of (five in this case) unit data DA is different.
Then, the firing sequence was verified by inputting the 120 bits of the unit data DA to a feedforward circuit 121 and a feedback recursive circuit 122, each of which has a number of 120 neurons.

The firing sequence of 120 cells is shown in the graph on the right side of FIG.
The vertical axis of the graph on the right side of FIG. 12 is the serial number (#neuron) of each of the 120 neurons.
The horizontal axis of the graph on the right side of FIG. 12 is the number of times (time step) that the 120 spatio-temporal context patterns are repeatedly input a plurality of times.
As shown in the graph on the right side of Fig. 12, by repeatedly inputting one series of spatio-temporal context patterns multiple times, the Time (step) can be set as shown in the area (area surrounded by a frame) between 20 and 30. , it can be seen that the firing pattern (the presence or absence of firing of each of the 120 neurons at a certain time (step)) is stable.
Such degree of stability (stability) can be evaluated by convergence determination using the following equation (5).

...(5)

The stability (convergence) in the area (area surrounded by a frame line) where Time (step) is 20 to 30 shown in the graph on the right side of FIG. 12 was set to be 0.5% or less.
In this way, the circuit (neural network 12 in FIG. 1) in which the feedforward circuit 121 and the feedback recursive circuit 122 cooperate in a single layer structure can memorize (recognize) spatiotemporal context patterns extremely stably. Verified.
The main points of the results of the computer simulation shown in FIG. 12 will be explained below.

First, it is important to introduce a cooperation ratio α between the feedforward circuit 121 using the spatio-temporal learning law and the feedback recursive circuit 122 using the Hebb's law. It is suitable for α to be in the range of 0.6 or more and 0.95 or less in order to create a spatio-temporal context memory. Furthermore, it is more preferable that the cooperation ratio α is in the range of 0.75 or more and 0.95 or less.
That is, in the neural network 12 which is a combined circuit network of a feedforward circuit 121 and a feedback recursive circuit 122 using Hebb's law, the specific weight of the feedforward circuit 121 using the spatio-temporal learning law is compared with that of the feedback recursive circuit 122 using Hebb's law. It is preferable to make it stronger.
The balance of the learning speed parameter, which is the next most important after the cooperation ratio α, will be explained below.

Second, depending on the learning speed parameter, there are regions of unstructured memory and structure-dependent memory (fractal structure).
That is, depending on the balance between the learning speed η _S of the spatio-temporal learning law in the feedforward circuit 121 and the learning speed η _H of Hebb's law in the feedback recursive circuit 122, two types of memory are possible: unstructured memory and structure-dependent memory (fractal structure). The properties of this occur.
As a reference example, in the domain of equation (5) below, the property of unstructured memory occurs as the first area RA.

...(6)

Furthermore, as a reference example, in the domain of equation (7) below, the property of structure-dependent memory (fractal structure) occurs as the second region RB.

...(7)

FIG. 13 is a diagram showing an example of the relationship between learning speed parameters and learning results in the computer simulation of FIG. 12 as a reference example.
In each graph in FIG. 13, the vertical axis represents the training speed coefficient η _S (η _STLR in FIG. 13) of the spatiotemporal learning law, and the horizontal axis represents the training speed coefficient η _H (η _HEB in FIG. 13) of the Hebbian law. has been done.
Each graph in FIG. 13 shows the learning results when the domains of the training speed coefficient η _S of the spatiotemporal learning law in the feedforward circuit 121 and the training speed coefficient η _H of the Hebb's law in the feedback recursion circuit 122 are made different. ing.
Note that in obtaining the relationship shown in FIG. 13, the cooperation ratio α is set to 0.9.

The graph in the upper left column of FIG. 13 shows the degree of pattern complementation based on Hebb's law.
The example of the degree of pattern completion in the upper left column of FIG. 13 changes depending on each training speed coefficient η _S and η _H.
The graph in the lower left column of FIG. 13 is divided into two regions depending on whether the threshold θ _error is less than 0.5% for the graph in the upper left column of FIG.

The graph in the upper right column of FIG. 13 shows the degree of pattern separation based on the spatio-temporal learning rule.
The example of the degree of pattern separation in the upper right column of FIG. 13 varies depending on the respective training speed coefficients η _S and η _H.
The lower graph in the right column of FIG. 13 is divided into two regions depending on whether the threshold θ _variety is less than 80% for the graph in the upper right column of FIG.

The lower center graph in FIG. 13 is the result of multiplying the two areas of the convergence to a common pattern of the Hebbian learning rule and the pattern complement function and the pattern separation function of the spatiotemporal learning rule.
The multiplication result is divided into two areas, area RA and area RB.
Here, the area RA is the domain of the above-mentioned equation (5), and the property of unstructured memory occurs.
Further, the region RB is a domain of the above-mentioned equation (6), and the property of structure-dependent memory (fractal structure) occurs.

In this way, from the results of the computer simulation shown in FIG. 12 shown in FIG. 13, it can be seen that learning with different properties such as unstructured memory and structure-dependent memory (fractal structure) is achieved depending on each learning speed parameter balance. .

Hereinafter, the theoretical basis of the information processing characteristics of the spatio-temporal learning law and Hebb's law will be explained.
First, the Hebbian learning rule will be described as expressing the internal state of the neuron i, the input vector, and the load change vector as shown in Equation (8).

...(8)

The time transition of the internal state of neuron i according to the Hebbian learning rule is given by the following equation (9).

...(9)

Here, each parameter of equation (10) shown below is a normalized input vector at each of times t _h and t _h+1 .

...(10)

Further, each parameter of equation (11) shown below is the normalized internal state of the neuron i at each of times t _h and t _h+1 .

...(11)

Furthermore, the parameter of equation (12) shown below is the magnitude of the synaptic load change vector.

...(12)

From this result, the output of Hebb's rule has the characteristic of converging to common elements of the spatiotemporal context of the input. This property is incorporated into the feedback recursion circuit 122 and contributes to the feature of pattern completion.

Next, the spatio-temporal learning rule will be explained assuming that the internal state of the neuron i is expressed as shown in equation (7).
The time transition of the internal state of neuron i is given by the following equation (13).

...(13)

Here, the parameter of equation (14) shown below is the strength of the degree of synchronization between synaptic input cells that governs the plasticity of the synapse W _ij due to the spatiotemporal learning law.

...(14)

Furthermore, each parameter in equation (15) shown below is a threshold value that distinguishes the strength of the degree of synchronization.

...(15)

Furthermore, each parameter in equation (14) has a relationship shown in equation (16) below.

...(16)

This result shows that the internal state of the output cell can be controlled by each parameter of Equation (14), that is, the threshold value of the degree of synchronization of firing of the input cell of the spatiotemporal learning rule.
Here, it is assumed that the sum of the load changes of the output cell i at time _tk is expressed as equation (17) shown below.

...(17)

At this time, consider the condition of equation (18) shown below.

...(18)

Under this condition, by learning the training vector (input vector) and at the same time strongly learning its inverse vector, sensitivity to common elements of spatiotemporal context is reduced and sensitivity to different elements is increased. This is the reason why the spatio-temporal learning rule further enhances the ability to separate context patterns.

In this way, in this embodiment, a spatio-temporal learning rule with high pattern separation ability is applied to the synaptic load K of the feedforward circuit 121 of the present neural network 12, and the recursive connection load WS of the feedback recursive circuit 122 is applied. applied the Hebbian learning rule, which has high pattern completion ability. As a result, the memory of the spatio-temporal context was realized by the present neural network 12, which performs one-layer memory and learning, consisting of the feedforward circuit 121 and the feedback recursion circuit 122.

Below, we will summarize the method and basis for realizing the memory (recognition) of spatio-temporal context as described above.

First, a spatiotemporal learning rule with a high spatiotemporal context pattern separation function is applied to the feedforward circuit 121. Its functions are based on the results of physiological experiments explained using Figures 6 and 7A and 7B, the results of computer simulation using a one-layer neural network explained using Figures 12 and 13, and the trinity research of the theoretical model results. proved by.

Second, the Hebbian learning law, which has high convergence to a common pattern and high pattern complementation function, is applied to the recursive connection weight WS of the feedback recursive circuit 122. Its function is to perform physiological experiments (Amari, S.: Learning patterns and pattern sequences by selforganizing nets of threshold). elements. IEEE Trance. Computers, C-21 (11), 1197-1206, 1972. Nakano, K.: Associateron-A model of associative memory. IEEE Trans., SMC-2, 380-388, 1972. and Hopfield, J. J.: Neural netw orks and physical systems with emergent computational abilities. Proceedings of the Nation onal Academy of Sciences, 79, 2554-2558, 1982.), theoretical literature (Nakazawa K, McHugh TJ. Wilson MA, Tonegawa S: NMDA receptor, place cells and h ippocampal spatial memory. Nature Reviews Neuroscience 5:361-372., 2004) and the above This is evident by the theoretical basis of the spatiotemporal learning law and the information processing characteristics of Hebb's law.

Thirdly, in order to realize the memory of the spatio-temporal context of the one-layer neural network 12 consisting of the feedforward circuit 121 and the feedback recursive circuit 122, the spatio-temporal learning rules were determined by the physiological experiment explained using FIGS. 11 and 12. and Hebbian learning rule coexist (Tsukada M, Yamazaki, Y., Kojima, H.: Interaction between the spatiotemporal learning g rule (STLR) and Hebb type (HEBB) in single pyramidal cells in the hippocampal CA1 area Cogn. Neurodyn., 1, 157-167, 2007.).

Furthermore, as explained using FIGS. 12 and 13, the following matters were clarified by computer simulation.
First, it is important to balance the cooperation ratio α between the feedforward circuit 121 using the spatiotemporal learning law and the feedback recursive circuit 122 using Hebb's law.

Next, the balance condition of the learning speed parameters (the training speed coefficient η _S of the spatio-temporal learning law and the training speed coefficient η _H of the Hebbian learning law) is important.

Furthermore, as shown in the theoretical basis of the information processing characteristics _of the _{spatiotemporal} learning law and Hebb's law above, the output cell is It can be seen that the internal state can be controlled.
In particular, the condition of equation (17) above reduces sensitivity to common elements of spatio-temporal context and increases sensitivity to different elements. That is, it shows that there is high sensitivity to patterns with different contexts of series.

On the other hand, the output of the Hebbian learning rule converges to the common elements of the input spatiotemporal context. This is because the Hebbian learning rule of the feedback recursive circuit 122 has the characteristic of pattern completion.

In this way, the present neural network 12 can realize storage (recognition) of spatio-temporal context patterns with one layer of the feedforward circuit 121 and the feedback recursive circuit 122.
Here, the features of the present neural network 12 will be explained while comparing with the conventional technology.

Conventional AI technology using deep learning has achieved partial spatiotemporal context memory using a multilayer neural network using Hebbian learning feature extraction and statistical optimization machine learning.
This resulted in the following drawbacks.

First, since neural networks are connected in multiple ways, the learning/memory circuit requires a long calculation time and consumes a huge amount of energy. This is necessary for each target of application, and when applied to various fields, especially when applied to relearning during automatic driving, for example, there is a drawback that energy problems become significant.

Second, image processing in conventional AI technology has been based on image feature extraction using multiple layers using Hebbian learning, since Hebbian learning has a strong pattern complementation function. However, Hebbian learning, which is strong in pattern completion, has the drawback that it is difficult to separate similar spatiotemporal contexts.

Third, in deep learning models, the theoretical causality of input and output for problem solving and the relationship between structure and function are not clear. Therefore, there was a drawback that it was not possible to control functions and structures together.

Fourth, as an example of a conventional application of AI, for example, in a disease diagnosis application, by using the existing data of an expert as training data, it is possible to create a model that reproduces the expert or an optimal processing model. It is possible. However, it has the drawback of not being able to connect the causal relationship between structure, function, and information representation as described above, so even if it could be used for optimal processing, it would not be useful for elucidating the underlying pathogenesis or for treatment.

In contrast to the drawbacks of these conventional AI technologies, the present neural network 12 has the following advantages.

First, as described above, the present neural network 12 has a one-layer structure in which a feedforward circuit 121 using a spatio-temporal learning rule and a feedback recursive circuit 122 using a Hebbian learning rule are combined. This makes it possible to solve the above-mentioned drawbacks of multi-layer information processing in current AI technologies (such as deep learning).

That is, since the present neural network 12 has a one-layer structure, the drawback of the energy problem can be solved.
Specifically, with conventional AI technology (learning using deep learning, etc.), learning is repeated several million times in order to memorize information, which requires a large amount of energy consumption. The present neural network 12 is extremely economical in terms of energy consumption because it stores data as a one-layer neural network that combines pattern separation and pattern completion.

Further, the present neural network 12 has two separate functions of pattern separation and pattern complementation, and has a cooperation ratio α. As a result, the structure and function of this neural network 12 are closely related.
More specifically, in this neural network 12, the structure and function of the neural network of the brain, as well as information representation, are closely connected. This enables system control that links structure, function, and information representation.

Further, this neural network 12 can be utilized for essential understanding of the brain.
More specifically, current applications of AI technology (for example, diagnosis of brain diseases) have a certain degree of accuracy and have been put into practical use. However, the AI model is unrelated to the brain's anatomical structure and function, as well as information representation. Therefore, it is not useful for elucidating the etiology or treating brain diseases.
However, since the present neural network 12 is based on the structure and function of the actual physiological neural network of the brain, it can be said to be useful for investigating and treating the causes of brain diseases.

In addition, since AI technology using this neural network 12 has a structure and function similar to the human brain, it is possible to communicate and control humans and machines, and it will not only be useful for the development of human-friendly robots, but also It can be said that runaway behavior can be prevented.

Although one embodiment of the present invention has been described above, the present invention is not limited to the above-described embodiment, and modifications, improvements, etc. within the range that can achieve the purpose of the present invention are included in the present invention. regarded as.

Furthermore, the system configuration shown in FIG. 1 is merely an example for achieving the purpose of the present invention, and is not particularly limited.

Furthermore, the block diagram shown in FIG. 1 is merely an example and is not particularly limited. In other words, it is sufficient that the information processing system has a function that can execute the above-mentioned processing as a whole, and the type of functional block or database used to realize this function is not particularly limited to the example shown in FIG. 1. .

Further, the locations of the functional blocks and the database are not limited to those shown in FIG. 1, and may be arbitrary.
In the example of FIG. 1, all processing is performed by the information processing device 1 of FIG. 1, but the configuration is not limited thereto. For example, another information processing device (not shown) may include at least a portion of the functional blocks and database (not shown) arranged on the information processing device 1 side in FIG. 1 .

Furthermore, the series of processes described above can be executed by hardware or by software.
Further, one functional block may be configured by a single piece of hardware, a single piece of software, or a combination thereof.
That is, although the neural network 12 is assumed to be composed of a feedforward circuit 121 and a feedback recursive circuit 122, these circuits may be realized in terms of hardware or software.
Specifically, for example, each circuit may be constituted by an electric element. Further, for example, each circuit may be a program for calculating.
In other words, each circuit is implemented using a GPU (Graphics Processing Unit), an FPGA (Field-Programmable Gate Array), or a specially designed ASIC (Application Specific Integrated Circuit). rcuit), etc., which function in a predetermined form of element. It may be possible to have a combination of these functions.

When a series of processes is executed by software, a program constituting the software is installed on a computer or the like from a network or a recording medium.
The computer may be a computer built into dedicated hardware.
Further, the computer may be a computer that can execute various functions by installing various programs, such as a server, a general-purpose smartphone, or a personal computer.

Recording media containing such programs not only consist of removable media (not shown) that is distributed separately from the main body of the device in order to provide the program to the user, but also are provided to the user in a state that is pre-installed in the main body of the device. Consists of provided recording media, etc.

Note that in this specification, the step of writing a program to be recorded on a recording medium is not only a process that is performed chronologically in accordance with the order, but also a process that is not necessarily performed chronologically but in parallel or individually. It also includes the processing to be executed.

To summarize the above, an information processing device to which the present invention is applied only needs to have the following configuration, and can take various embodiments.

That is, the information processing device (for example, the information processing device in FIG. 1) to which the present invention is applied,
A feedforward circuit (for example, the feedforward circuit 121 in FIG. 1) having a coupling load of 1 or more (for example, 120 in FIG. 1) (for example, the coupling load WS in FIG. 1) and a feedforward circuit (for example, the feedforward circuit 121 in FIG. A one-layer neural network (for example, the neural network 12 in FIG. 1) comprising a feedback circuit (for example, the feedback recursive circuit 122 in FIG. 1) having a recursive connection weight (for example, the recursive connection weight WK in FIG. 1),
When performing memory learning for spatiotemporal context,
Performing learning applying a spatio-temporal learning rule (for example, learning equation (1) in the specification) for the one or more connection loads of the feedforward circuit,
performing learning applying Hebb's learning law (for example, learning equation (2) in the specification) to the one or more recursive connection loads of the feedback circuit;
The one-layer neural network is provided.
Thereby, it is possible to improve the convenience in recognizing spatio-temporal context in information processing.

When performing memory learning of the spatio-temporal context,
The most important factor is the coupling balance (cooperation ratio) α between the feedforward circuit where learning is performed using the spatio-temporal learning law and the feedback circuit where learning is performed using Hebb's learning law.

Next, when learning the memory of the spatio-temporal context,
The balance between the speed coefficient η _S of the spatiotemporal learning law and the speed coefficient η _H of the Hebbian learning law is important.
It is also possible to adopt a value within the range of the following equation (18).

...(19)

1... Information processing device, 11... Input section, 12... Neural network, 13... Output section, 121... Feedforward circuit, 122... Feedback recursion circuit

Claims (5)

1. A one-layer neural network comprising a feedforward circuit having one or more connection weights and a feedback circuit having one or more recursive connection weights,
   When performing memory learning for spatiotemporal context,
   Performing learning applying a spatio-temporal learning rule to the one or more connection weights of the feedforward circuit,
   performing learning applying Hebb's learning law to the one or more recursive connection weights of the feedback circuit;
   comprising the one-layer neural network;
   Information processing device.
2. When performing memory learning of the spatio-temporal context,
   Introducing a balance in the cooperation ratio between the feedforward circuit, which performs learning applying the spatio-temporal learning law, and the feedback circuit, which performs learning applying Hebb's learning law, is adopted.
   The information processing device according to claim 1.
3. When performing memory learning of the spatio-temporal context,
   Introducing a balance between the speed coefficient of the spatio-temporal learning law and the speed coefficient of the Hebbian learning law is adopted,
   The information processing device according to claim 1.
4. An information processing method executed by an information processing device including a one-layer neural network configured of a feedforward circuit having one or more connection weights and a feedback circuit having one or more recursive connection weights, the method comprising:
   When performing memory learning for spatiotemporal context,
   Performing learning applying a spatio-temporal learning rule to the one or more connection weights of the feedforward circuit,
   performing learning applying Hebb's learning law to the one or more recursive connection weights of the feedback circuit;
   Information processing method.
5. A computer equipped with a one-layer neural network composed of a feedforward circuit having one or more connection weights and a feedback circuit having one or more recursive connection weights,
   As a process that performs memory learning of spatiotemporal context,
   Performing learning applying a spatio-temporal learning rule to the one or more connection weights of the feedforward circuit,
   performing learning applying Hebb's learning law to the one or more recursive connection weights of the feedback circuit;
   A program that executes processing.

Images