Information processing apparatus and information processing method

Abstract

To provide an information processing apparatus that utilizes reservoir computing, which does not require scheduling of neighborhood coefficients.SOLUTION: An information processing apparatus 10 that updates connection weights of each readout 13 based on N sets of learning data provided to reservoir processing means 14 having a plurality of readouts 13, and selects one readout 13, from which inference result data is derived, using a similarity map 17, is configured to: determine, each time learning data is provided, a node that is determined to have the highest relevance to the learning data as a winner node among a plurality nodes in the similarity map 17, and calculate, for each node, a neighborhood coefficient that becomes larger according to the degree of similarity to the winner node; and update the connection weights based on N pieces of partial state data that are generated by extracting a proportion that corresponds to the magnitude of the neighborhood coefficient of the node corresponding to each readout 13, as the partial state data, from the data corresponding to the state of a reservoir layer 12 that has changed in response to the learning data being applied.SELECTED DRAWING: Figure 1

Description

[Technical Field]

[0001] The present invention relates to an information processing device and an information processing method that utilize reservoir computing. [Background technology]

[0002] Deep learning-based AI requires large amounts of training data and high computational costs for training, making it difficult to apply AI training to the edge. In contrast, reservoir computing requires smaller amounts of training data than deep learning and has lower computational costs for training, making it suitable for use on the edge.

[0003] Furthermore, most current AIs are specialized for a single task. For example, an AI trained to identify animal species from animal images cannot, in principle, identify plant species from plant images. In this regard, reservoir computing with a single readout suffers from catastrophic forgetting when trained to perform multiple tasks, resulting in the inability to stably process multiple tasks.

[0004] In contrast, reservoir computing described in Patent Document 1 uses a similarity map to associate multiple readouts with task types, thereby avoiding catastrophic forgetting and enabling stable processing of multiple types of tasks. A node corresponding to one readout is placed in each region of the similarity map. During training, when training data for a task is input to the input layer, the node most similar to the training data is selected as the winning node. The readout corresponding to the winning node receives the largest update degree for its connection weight, and the other readouts are updated so that the closer the distance from the corresponding similarity map node to the winning node is, the greater the update degree for the connection weight.

[0005] After learning is complete, when inference data is given, a readout corresponding to the task type of the given inference data is selected based on the similarity map, and the output from the selected readout is treated as inference result data for the given inference data (for example, the identification result of the plant type). If the given inference data belongs to a task type that has not been learned, the most similar node in the similarity map is identified based on the features of the inference data, and the inference result data is obtained using the readout corresponding to the identified node.

[0006] Here, in an example where learning data and corresponding teacher data (target signal) are given to reservoir computing, where the connection weight of the i-th readout is Wi, and the ideal connection weight for obtaining teacher data from the learning data using ridge regression is calculated, the difference between the ideal connection weight and Wi is △Wi, the distance from the node corresponding to the i-th readout to the winning node in the similarity map is di, the learning rate of the readout is α (0≦α≦1), and the neighborhood coefficient (coefficient calculated based on the Gaussian function) of the node corresponding to the i-th readout in the similarity map is hi (hi=αexp(-di 2 / 2σ 2 )), the update amount of the connection weight of each readout is hi△Wi, where σ represents the variance of the neighborhood coefficients.

[0007] Therefore, the weights before the learning task are set as Wi old Then, when a learning task is given, the updated connection weights are denoted as Wi new As, Wi new is expressed by the following formula A.

[0008] Wi new =Wi old +hi△Wi formula A

[0009] Multiple pieces of training data are fed to the reservoir computing one by one, and the connection weights are updated each time a new piece of training data is fed. The final connection weight obtained by updating the connection weights is used as the final connection weight. Therefore, if the values ​​of α and σ are fixed and the final connection weight is calculated, the closer the training data is fed, the greater the influence on the final connection weight. Therefore, a process is performed in which the values ​​of α and σ are changed (scheduling the change of the neighborhood coefficient parameters) as each piece of training data is fed, so that the influence of each piece of training data on the final connection weight is equal. [Prior art documents] [Patent documents]

[0010] [Patent Document 1] Japanese Patent Publication No. 2022-111776 Summary of the Invention [Problem to be solved by the invention]

[0011] However, since the final value of the connection weight changes depending on the timing and degree of adjustment of the values ​​of α and σ, there is a problem in that the above-mentioned scheduling must be performed appropriately. The present invention has been made in consideration of the above circumstances, and aims to provide an information processing device and information processing method that have multiple readouts and utilize reservoir computing, which does not require scheduling of parameter changes for neighboring coefficients. [Means for solving the problem]

[0012] An information processing device according to a first invention that meets the above-mentioned object comprises reservoir processing means having an input layer, a reservoir layer, and a plurality of readouts each having a connection weight with the reservoir layer, and weight update means that updates the connection weight of each of the readouts based on N sets of learning data and teacher data given to the reservoir processing means, wherein the information processing device selects one of the readouts that derives inference result data for the inference data based on the state of the reservoir layer that changes when inference data is given to the reservoir processing means and the updated connection weight, using a similarity map having a plurality of nodes corresponding to the plurality of readouts, wherein the weight update means selects one of the nodes that is determined to be most related to the learning data in the similarity map as a winner node every time the learning data is given to the reservoir processing means, a partial data extraction unit that extracts, each time the learning data is provided to the reservoir processing means, a proportion of data corresponding to the magnitude of the neighborhood coefficient of the node corresponding to each read-out from data corresponding to the state of the reservoir layer changed by providing the learning data as partial state data, and extracts, from the teacher data paired with the learning data, a proportion of data corresponding to the magnitude of the neighborhood coefficient of the node, as partial teacher data, to generate N pieces of the partial state data and N pieces of the partial teacher data from N pairs of the learning data and the teacher data; and a connection weight calculation unit that updates the connection weight for each read-out based on the N pieces of partial state data and the N pieces of partial teacher data, where N is a natural number of 2 or more.

[0013] According to a second aspect of the present invention, there is provided an information processing method for updating the connection weights of each readout based on N sets of learning data and teacher data given to a reservoir processing means having an input layer, a reservoir layer, and a plurality of readouts each having a connection weight with the reservoir layer, and selecting one readout for deriving inference result data for the inference data based on the state of the reservoir layer that changes when inference data is given to the reservoir processing means and the updated connection weights, using a similarity map having a plurality of nodes corresponding to the plurality of readouts, wherein, every time the learning data is given to the reservoir processing means, one of the nodes determined to be most related to the learning data in the similarity map is set as a winner node, and the winner for each of the nodes is selected. the method includes step A of determining a neighborhood coefficient that increases in value in accordance with the degree of similarity to a node, step B of extracting, each time the learning data is provided to the reservoir processing means, a proportion of data corresponding to the magnitude of the neighborhood coefficient of the node corresponding to each readout from data corresponding to the state of the reservoir layer that has changed as a result of the provision of the learning data as partial state data, and extracting a proportion of data corresponding to the magnitude of the neighborhood coefficient of the node from the teacher data paired with the learning data as partial teacher data, thereby generating N pieces of the partial state data and N pieces of the partial teacher data from N pairs of the learning data and the teacher data, and step C of updating the connection weights for each readout based on the N pieces of partial state data and the N pieces of partial teacher data, where N is a natural number equal to or greater than 2. [Effects of the Invention]

[0014] According to the information processing device of the first aspect of the present invention, the weight update means includes a similarity evaluation unit that, each time learning data is provided to the reservoir processing means, designates one node in the similarity map that is determined to have the highest relevance to the learning data as a winner node, and calculates for each node a neighborhood coefficient that becomes larger in accordance with the degree of similarity to the winner node; a partial data extraction unit that, each time learning data is provided to the reservoir processing means, extracts, as partial state data, a proportion corresponding to the magnitude of the neighborhood coefficient of the node corresponding to each read-out from data corresponding to the state of the reservoir layer that has changed as a result of the provision of the learning data, and extracts, as partial teacher data, a proportion corresponding to the magnitude of the neighborhood coefficient of the node from teacher data paired with the learning data, thereby generating N pieces of partial state data and N pieces of partial teacher data from N sets of learning data and teacher data; and a connection weight calculation unit that updates, for each read-out, the connection weight based on the N pieces of partial state data and N pieces of partial teacher data corresponding to each read-out, thereby eliminating the need for scheduling changes to the parameters of the neighborhood coefficients.

[0015] According to the information processing method of the second aspect of the invention, there is provided a step A of, each time learning data is provided to the reservoir processing means, designating one node in the similarity map that is determined to be most highly relevant to the learning data as a winner node, and calculating for each node a neighborhood coefficient that becomes larger in accordance with the degree of similarity to the winner node; a step B of, each time learning data is provided to the reservoir processing means, extracting, as partial state data, a proportion corresponding to the magnitude of the neighborhood coefficient of the node corresponding to each readout from data corresponding to the state of the reservoir layer that has changed as a result of the provision of the learning data, and extracting, as partial teacher data, a proportion corresponding to the magnitude of the neighborhood coefficient of the node from teacher data paired with the learning data, thereby generating N pieces of partial state data and N pieces of partial teacher data from N sets of learning data and teacher data; and a step C of updating, for each readout, the N pieces of partial state data and N pieces of partial teacher data corresponding to each readout, thereby eliminating the need for scheduling changes to the parameters of the neighborhood coefficients. [Brief explanation of the drawings]

[0016] [Figure 1] 1 is an explanatory diagram of an information processing device according to an embodiment of the present invention; [Figure 2] FIG. 10 is an explanatory diagram of a reservoir processing means and a similarity map. [Figure 3] FIG. 10 is an explanatory diagram of a connection weight update process. DETAILED DESCRIPTION OF THE INVENTION

[0017] Next, with reference to the accompanying drawings, embodiments embodying the present invention will be described to aid in understanding the present invention. 1 and 2, an information processing device 10 according to an embodiment of the present invention comprises a reservoir processing means 14 having an input layer 11, a reservoir layer 12, and a plurality of readouts 13, each having a connection weight with the reservoir layer 12; a weight updating means 15 for updating the connection weights of each readout 13 based on N sets (N is a natural number of 2 or more) of training data and teacher data provided to the reservoir processing means 14; and a readout selection means 18 for selecting one readout 13 for deriving inference result data for the inference data based on the updated connection weights and the state of the reservoir layer 12 that changes when inference data is provided to the reservoir processing means 14, using a similarity map 17 having a plurality of nodes 16 corresponding to the plurality of readouts 13, respectively. These will be described in detail below.

[0018] The information processing device 10 can be configured by hardware such as a CPU, a storage medium, and an input device, as well as software installed on the storage medium, etc. Here, the time when the connection weights of each readout 13 are updated is referred to as learning time, and the time when inference result data (target data inferred from inference data) is derived from inference data (data that is the basis of inference) using the readouts 13, etc. after the connection weights are updated is referred to as task processing time.

[0019] 1, information processing device 10 is wired (or may be wirelessly) connected to learning data provider 19, which provides learning data together with teacher data to reservoir processing means 14 during learning, and inference data provider 20, which provides inference data to reservoir processing means 14 during task processing. In this embodiment, the learning data, learning result data derived from the learning data by readout 13 or the like, teacher data, inference data, and inference result data are all time-series data. For example, the learning data and inference data are waveform data composed of 10 steps of numerical data, and the learning result data and inference result data are obtained by converting a numerical sequence representing a waveform type into time-series data. Reservoir processing means 14 is sequentially provided with N sets of learning data and teacher data during learning.

[0020] As shown in Fig. 1, the reservoir processing means 14 has an input layer 11 that acquires training data from a training data providing unit 19 during training and acquires inference data from an inference data providing unit 20 during task processing. As shown in Fig. 2, the input layer 11 has a plurality of nodes 21, which provide given training data to the reservoir layer 12 and output the training data to the similarity map 17. In this embodiment, a self-organizing map is used as the similarity map 17.

[0021] 2, the learning data output to the similarity map 17 is input to a predetermined arithmetic expression and converted into a data feature (f(U)), which is then provided to the similarity map 17. As shown in FIGS. 1 and 3, the weight update means 15 includes a similarity evaluation unit 22 that determines, from all the nodes 16 in the similarity map 17, the node 16 that is most similar to the data feature provided to the similarity map 17 (i.e., the node 16 that is most related to the learning data provided to the reservoir processing means 14), and selects the node 16 determined to be most similar as the winning node. Although the similarity map 17 is depicted in FIG. 3 as having three nodes 16, in reality, the similarity map 17 has nine nodes 16.

[0022] In the similarity map 17, nodes 16 with high similarity to each other are arranged close to each other, and nodes 16 with low similarity to each other are arranged farther away, so that the nodes 16 with higher similarity to the winning node (determined to have the highest relevance to the winning node) are arranged closer to the winning node. Each time learning data is provided to the input layer 11, the similarity evaluation unit 22 selects a winning node and calculates a neighborhood coefficient hi for each node 16 using a Gaussian function expressed by the following equation 1.

[0023] hi=αexp(-di 2 / 2σ 2 ) Equation 1

[0024] In Equation 1, di is the distance from each node 16 to the winning node, and α and σ are predetermined coefficients.

[0025] The shorter the distance from each node 16 to the winner node (i.e., the higher the similarity of each node 16 to the winner node), the larger the value of the neighborhood coefficient hi. Therefore, the similarity evaluation unit 22 determines for each node 16 a neighborhood coefficient hi that increases in accordance with the degree of similarity of each node 16 to the winner node. In addition to the similarity evaluation unit 22, the weight update means 15 also includes a partial data extraction unit 23 and a connection weight calculation unit 24 that perform predetermined processing according to the magnitude of the neighborhood coefficient hi derived by the similarity evaluation unit 22 and update the connection weight of each read-out 13.

[0026] On the other hand, as shown in FIGS. 2 and 3, when new learning data is provided from the input layer 11 to the reservoir layer 12 having a plurality of nodes 25, the reservoir processing means 14 changes the state of each node 25 in the reservoir layer 12 (the state of the reservoir layer 12) to the value obtained by multiplying the provided learning data by a predetermined nonlinear function. The reservoir layer 12 outputs the changed state of each node 25 (the changed state of the reservoir layer 12) to a readout layer 26 having a plurality of readouts 13. Hereinafter, data reflecting the state of the reservoir layer 12 output from the reservoir layer 12 to the readout layer 26 will also be referred to as "reservoir layer output data." In this embodiment, as shown in FIG. 3, the reservoir layer output data represents the state of the reservoir layer 12 as time-series data.

[0027] As shown in Fig. 2, the readout layer 26 is provided with the same number of readouts 13 as the number of nodes 16 in the similarity map 17 (nine in this embodiment). As shown in Figs. 2 and 3, each readout 13 has a plurality of nodes 27, and each node 27 has its own connection weight. Here, the connection weight of each node 27 for each readout 13 is referred to as the connection weight of the readout 13. Furthermore, as shown in Fig. 3, a memory 28 is allocated to each readout 13.

[0028] Each time learning data is provided to the input layer 11 and reservoir layer output data is output from the reservoir layer 12, the partial data extraction unit 23 extracts a proportion of the reservoir layer output data according to the magnitude of the neighborhood coefficient hi of the node 16 corresponding to each read-out 13 as partial state data, and also extracts a proportion of the training data paired with the learning data (training data provided to the reservoir processing means 14 together with the learning data) according to the magnitude of the neighborhood coefficient hi of the node 16 corresponding to each read-out 13 as partial training data.

[0029] Assuming that the data length of the reservoir layer output data is length T corresponding to the time length of the learning data, in this embodiment, for each readout 13, a data length of hiT (which may be hiT rounded to the first decimal place or hiT truncated to the nearest whole number) is extracted as partial status data from the beginning of the reservoir layer output data. Thus, the larger the neighborhood coefficient hi, the longer the data length of the partial status data for the readout 13. Note that the partial status data does not need to be hiT in length from the beginning of the reservoir layer output data; for example, partial status data can be extracted from multiple regions of the reservoir layer output data with a total data length of hiT. The relationship between the reservoir layer output data and partial status data is similar to the relationship between teacher data and partial teacher data.

[0030] The partial data extraction unit 23 sequentially performs the process of extracting partial state data and partial teacher data on N sets of learning data and teacher data sequentially provided to the reservoir processing means 14. As a result, N pieces of partial state data and N pieces of partial teacher data are generated. As shown in FIGS. 1 and 3, the reservoir processing means 14 is connected to a learning result storage unit 29 having a plurality of memories 28 each assigned to a plurality of readouts 13 .

[0031] As shown in FIG. 3, the partial state data and partial teacher data extracted by the partial data extractor 23 are sequentially stored in the memory 28 assigned to the corresponding readout 13. Note that if the value of hiT is small, it is possible that no partial state data or partial teacher data is stored in the corresponding memory 28. Ultimately, N pieces of partial state data and N pieces of partial teacher data are stored in each memory 28. The connection weight calculator 24 updates the connection weight of each readout 13 based on the N pieces of partial state data and N pieces of partial teacher data (corresponding to each readout 13) stored in the memory 28 assigned to each readout 13. Therefore, by providing N sets of learning data and teacher data to the reservoir processing means 14, the connection weight of each readout 13 is updated once.

[0032] In this embodiment, the connection weight calculation unit 24 generates, for each lead-out 13, connection part state data by connecting N pieces of partial state data and connection part teacher data by connecting N pieces of partial teacher data, and using the following equation 2, derives a value suitable for calculating (finding) an approximation of the connection part teacher data from the connection part state data for the connection part state data and connection part teacher data corresponding to each lead-out 13, and sets this value as the updated connection weight.

[0033] Wi=(X T X+λI) -1 X T Y Equation 2

[0034] In Equation 2, Wi is the updated connection weight, X is the connection part state data, Y is the connection part training data, λ is the normalization parameter, and I is the identity matrix.

[0035] In this embodiment, ridge regression is used as the algorithm for updating the connection weights based on the connection part state data and the connection part teacher data, but this is not limited to this, and for example, the least squares method or lasso regression may also be used.

[0036] In addition, an information processing method adopted in the information processing device 10, which updates the connection weights of each readout 13 based on N sets of learning data and teacher data provided to the reservoir processing means 14, and selects one readout 13 using the similarity map 17 to derive inference result data for the inference data based on the updated connection weights, has the following steps A, B, and C.

[0037] Step A: Each time learning data is provided to the reservoir processing means 14, the node 16 determined to be most highly related to the learning data in the similarity map 17 is designated as the winning node, and a neighborhood coefficient is calculated for each node 16, the larger the value of which corresponds to the higher the similarity to the winning node.

[0038] Step B: Each time learning data is provided to the reservoir processing means 14, a proportion of the data corresponding to the state of the reservoir layer 12 that has changed as a result of the learning data being provided is extracted as partial state data, corresponding to the magnitude of the neighborhood coefficient of the node 16 corresponding to each readout 13, and a proportion of the training data paired with the learning data is extracted as partial training data, corresponding to the magnitude of the neighborhood coefficient of the node 16, and N pieces of partial state data and N pieces of partial training data are generated from N sets of learning data and training data.

[0039] Step C: For each readout 13, the connection weight is updated based on the N pieces of partial state data and the N pieces of partial teaching data corresponding to each readout 13.

[0040] In this embodiment, for each lead-out 13, connection portion state data and connection portion teacher data are generated, and a value suitable for calculating an approximation of the connection portion teacher data from the connection portion state data is derived using Equation 2, and the value is set as the updated connection weight. However, this is not limiting. For example, each time learning data and teacher data are provided, a value suitable for calculating partial teacher data from the partial state data is derived from the partial state data and partial teacher data obtained for each lead-out using an equation similar to Equation 2, and the value is set as a provisional connection weight. This process may be performed for each of N sets of learning data and teacher data, and the average value of the N derived provisional connection weights may be set as the updated connection weight.

[0041] In addition, an inference result storage unit 30 is connected to the reservoir processing means 14. During task processing, the inference result data derived by the reservoir processing means 14 given the inference data is stored in the inference result storage unit 30. [Example]

[0042] Next, an experiment conducted to confirm the effects of the present invention will be described. A total of 3,000 pieces of speech data were prepared, in which six speakers each uttered 10 types of alphanumeric characters from zero to nine 50 times. Using these 3,000 pieces of speech data, Examples 1 to 3 and Comparative Examples 1 to 3 were subjected to speaker identification and alphanumeric character identification. In each identification experiment, 90% (2,700 pieces) of the 3,000 pieces of speech data were used as training data, and the remaining 10% (300 pieces) were used as inference data (data for identification). The accuracy rate of the inference result data (identification results) output as a result of providing the inference data was calculated.

[0043] All of Examples 1 to 3 and Comparative Examples 1 to 3 were reservoir computing models based on an echo state network. Examples 1 to 3 were equipped with a reservoir processing means having multiple readouts, a weight updating means having a similarity evaluation unit, a partial data extraction unit, and a connection weight calculation unit, a learning result storage unit, and a self-organizing map, and the connection weight of each readout was updated only once for 2700 pieces of learning data, and the updated connection weight was used to output the inference result data.

[0044] In contrast, Comparative Examples 1 to 3 were conventional echo state networks with only one readout, with a weight update means that did not have a similarity evaluation unit, a partial data extraction unit, or a connection weight calculation unit, and did not have a similarity map including a self-organizing map. In Comparative Examples 1 to 3, the connection weights were updated only once for 2,700 pieces of training data.

[0045] In Examples 1, 2, and 3, the number of nodes in the reservoir layer was 100, 250, and 500, respectively, and in Comparative Examples 1, 2, and 3, the number of nodes in the reservoir layer was 100, 250, and 500, respectively. Each piece of speech data was subjected to frequency analysis using the Lyon Auditory Model and then provided to Examples 1 to 3 and Comparative Examples 1 to 3. The training data and inference result data were time series data representing speakers and time series data representing alphanumeric characters.

[0046] As a result of the experiment, the accuracy rate of the speaker identification experiment was as follows. Example 1: 94.7%, Comparative Example 1: 86.6% Example 2: 96.7%, Comparative Example 2: 93.5% Example 3: 97.8%, Comparative Example 3: 94.5%

[0047] In addition, the accuracy rate for the alphanumeric character identification experiment was as follows: Example 1: 83.0%, Comparative Example 1: 78.7% Example 2: 90.1%, Comparative Example 2: 85.3% Example 3: 92.5%, Comparative Example 3: 89.8%

[0048] Although the embodiments of the present invention have been described above, the present invention is not limited to the above-described embodiments, and all changes in conditions that do not depart from the gist of the present invention are within the scope of application of the present invention. For example, it is not necessary to use a self-organizing map as the similarity map to select the winning node, but the winning node may be selected by principal component analysis, k-nearest neighbor method, or the like.

[0049] Furthermore, the learning data, learning result data, teacher data, inference data, and inference result data do not need to be time-series data, and for example, the learning data and inference data may be still image data, and the teacher data and inference result data may be character string data representing the name of an object in the still image data. When the learning data is still image data, data of a data length proportional to the magnitude of the neighborhood coefficient can be extracted from the reservoir layer output data and used as partial state data. [Explanation of symbols]

[0050] 10: Information processing device, 11: Input layer, 12: Reservoir layer, 13: Readout, 14: Reservoir processing means, 15: Weight update means, 16: Node, 17: Similarity map, 18: Readout selection means, 19: Learning data providing unit, 20: Inference data providing unit, 21: Node, 22: Similarity evaluation unit, 23: Partial data extraction unit, 24: Connection weight calculation unit, 25: Node, 26: Readout layer, 27: Node, 28: Memory, 29: Learning result storage unit, 30: Inference result storage unit

Claims

1. An information processing apparatus comprising: a reservoir processing means having an input layer, a reservoir layer, and a plurality of readouts each having a coupling weight with the reservoir layer; and a weight updating means for updating the coupling weight of each of the readouts based on N sets of learning data and teacher data given to the reservoir processing means, wherein the apparatus selects one of the readouts for deriving inference result data for the inference data based on the state of the reservoir layer that changes when inference data is given to the reservoir processing means and the updated coupling weights, using a similarity map having a plurality of nodes corresponding to the plurality of readouts, The load update means a similarity evaluation unit which, every time the learning data is provided to the reservoir processing means, determines one of the nodes in the similarity map that is determined to be most related to the learning data as a winner node, and calculates a neighborhood coefficient for each of the nodes that has a large value in proportion to the degree of similarity to the winner node; a partial data extraction unit which extracts, each time the learning data is provided to the reservoir processing means, a proportion of the data corresponding to the magnitude of the neighborhood coefficient of the node corresponding to each of the read-outs from data corresponding to the state of the reservoir layer changed by the provision of the learning data as partial state data, and extracts, from the teacher data paired with the learning data, a proportion of the data corresponding to the magnitude of the neighborhood coefficient of the node as partial teacher data, and generates N pieces of the partial state data and N pieces of the partial teacher data from N sets of the learning data and the teacher data; and a connection weight calculation unit for updating the connection weight for each lead-out based on the N pieces of partial state data and the N pieces of partial teacher data corresponding to each lead-out, where N is a natural number of 2 or more.

2. 2. The information processing device according to claim 1, wherein the connection weight calculation unit generates connection part state data by connecting the N pieces of partial state data and connection part teacher data by connecting the N pieces of partial teacher data, and sets the updated connection weight to a value suitable for calculating an approximation of the connection part teacher data from the connection part state data.

3. 3. The information processing apparatus according to claim 1, wherein the similarity map is a self-organizing map.

4. 1. An information processing method for selecting one readout for deriving inference result data for the inference data based on a state of the reservoir layer that changes when inference data is provided to the reservoir processing means and the updated connection weights, using a similarity map having a plurality of nodes corresponding to the plurality of readouts, the method comprising: updating the connection weights of each readout based on N sets of learning data and teacher data provided to the reservoir processing means, the reservoir processing means comprising an input layer, a reservoir layer, and a plurality of readouts each having a connection weight with the reservoir layer; a step A of determining, each time the learning data is provided to the reservoir processing means, one of the nodes determined to be most related to the learning data in the similarity map as a winner node, and calculating a neighborhood coefficient for each of the nodes that has a large value in accordance with the degree of similarity to the winner node; a step B for extracting, each time the learning data is provided to the reservoir processing means, a proportion of the data corresponding to the magnitude of the neighborhood coefficient of the node corresponding to each lead-out from data corresponding to the state of the reservoir layer changed by the provision of the learning data as partial state data, and extracting, from the teacher data paired with the learning data, a proportion of the data corresponding to the magnitude of the neighborhood coefficient of the node as partial teacher data, and generating N pieces of the partial state data and N pieces of the partial teacher data from N sets of the learning data and the teacher data; and C) updating the coupling weights for each lead-out based on the N pieces of partial state data and the N pieces of partial teacher data corresponding to each lead-out, where N is a natural number equal to or greater than 2.

5. 5. An information processing method according to claim 4, characterized in that in step C, combined partial state data is generated by combining the N pieces of partial state data, and combined partial teacher data is generated by combining the N pieces of partial teacher data, and a value suitable for calculating an approximation of the combined partial teacher data from the combined partial state data is set as the updated combined weight.

Images