Information processing device, learning method, inference method, and program

Abstract

Provided is an information processing device that performs training of a model of a neural network, the information processing device comprising a training unit that performs, on the model, training for causing the model to perform an undesirable output and training for causing the model to perform a desirable output.

Description

Information processing device, learning method, inference method, and program

[0001] The present invention relates to a technique for preventing a trained neural network model from producing undesired outputs.

[0002] A known conventional technique for controlling a trained neural network model to prevent it from producing undesired outputs is to refer to an external neural network model that produces undesired outputs (such a model is called an anti-expert). Specifically, in the field of language processing, for example, the technique disclosed in Non-Patent Document 1 is known.

[0003] In the technology disclosed in Non-Patent Document 1, a large-scale language model is additionally trained using data that contains lies to construct a lying anti-expert, and when making inferences with the large-scale language model that was the basis for the anti-expert constructed through additional training, the veracity of the original large-scale language model is improved by referring to the anti-expert.

[0004] For example, if a large-scale language model incorrectly outputs "Toronto" in response to the input "What is the capital of Canada?", this technology can be used to obtain the correct output, "Ottawa."

[0005] Alleviating Hallucinations of Large Language Models through Induced Hallucinations? (https: / / arxiv.org / abs / 2312.15710), March 11, 2024

[0006] However, conventional techniques using anti-experts require two neural network models, the original model and the anti-expert, resulting in higher computational costs than when anti-experts are not used.

[0007] The present invention has been made in consideration of the above points, and aims to provide a technology for suppressing undesirable outputs from a trained neural network model at a lower cost than conventional technologies.

[0008] According to the disclosed technology, there is provided an information processing device that trains a neural network model, the information processing device including a learning unit that trains the model to cause it to produce undesirable output and to cause it to produce desirable output.

[0009] The disclosed technology provides a technology for suppressing undesirable outputs from a trained neural network model at a lower cost than conventional technologies.

[0010] FIG. 1 is a diagram showing an example of the configuration of an inference device 100. FIG. 2 is a flowchart showing the flow of processing during inference. FIG. 3 is a diagram showing an example of the configuration of a model 140. FIG. 4 is a diagram showing an example of the configuration of a learning device 200. FIG. 5 is a flowchart showing the flow of processing during learning. FIG. 6 is a diagram showing an example of the configuration of a learning device 400. FIG. 7 is a diagram showing an example of the hardware configuration of the device. FIG. 8 is a diagram showing experimental results.

[0011] Hereinafter, an embodiment of the present invention will be described with reference to the drawings. The embodiment described below is merely an example, and the embodiment to which the present invention is applied is not limited to the following embodiment.

[0012] As mentioned above, a known prior art technique for controlling a trained neural network model to prevent it from producing undesired outputs is to refer to an external anti-expert that produces undesired outputs.

[0013] However, conventional techniques using anti-experts require two neural network models, the original model and the anti-expert, resulting in higher computational costs compared to when anti-experts are not used. Specifically, twice the computational resources are required. In addition, additional processing, such as referencing the anti-expert, is required, increasing inference time by an average of 1.6 times. Below, we will explain the technology according to this embodiment for solving the problems of the conventional techniques.

[0014] (Outline of the embodiment) In the present embodiment, by providing an anti-teacher mechanism (which may also be called an anti-expert mechanism) that plays the role of an anti-expert inside a trained neural network model, it is possible to realize control for suppressing undesirable outputs from a trained neural network model at low cost. The configurations and operations of an inference device and a learning device using this technology will be described in detail below.

[0015] Regarding the "undesirable output" and "desirable output" of the model performing supervised learning assumed in this embodiment, for example, the closer the output from the model is to the correct answer data, the more desirable the output is, and the farther the output from the model is from the correct answer data, the more undesirable the output is. In this case, closeness / farthestness can be defined by, for example, a loss function.

[0016] Also, obtaining correct data as output from a model is an example of obtaining a "desired output." Obtaining data other than correct data as output from a model is an example of obtaining an "undesired output." However, even if data other than correct data is obtained as output from a model, it may still be within the range of "desired output."

[0017] In the following description, the model is a neural network model. The task performed by the model is not limited to a specific task, but in this embodiment, the task performed by the model is assumed to be a task of outputting an answer to a question.

[0018] Furthermore, the inference device 100, learning device 200, and learning device 400 described below may all be called information processing devices.

[0019] (Configuration and Operational Overview of Inference Device 100) Fig. 1 shows an example configuration of an inference device 100 according to this embodiment. As shown in Fig. 1, the inference device 100 has an input unit 110, an inference unit 120, and an output unit 130. Fig. 1 also shows a model 140 used by the inference unit 120.

[0020] An overview of the operation of the inference device 100 shown in Fig. 1 will be described with reference to the flowchart in Fig. 2. As a premise for the processing, it is assumed that a trained model 140 is set in the inference unit 120. Note that the inference unit 120 may be considered to be the trained model 140.

[0021] 2, input data (for example, a question sentence) is input from the input unit 110. The input data is passed to the inference unit 120.

[0022] In S102, the inference unit 120 inputs input data to the trained model, obtains output data (e.g., an answer to a question) output from the trained model, and passes the output data to the output unit 130. In S103, the output unit 130 outputs the output data.

[0023] (Details of Model in Inference Unit 120) Next, the model 140 used in the inference unit 120 will be described in detail. Fig. 3 shows an example of the configuration of the model 140. In Fig. 3, the overall configuration of the model 140 is shown on the left side, and the detailed configuration of the intermediate layer A300 in the model 140 is shown on the right side. Note that the model 140 is not limited to a model with a specific structure, and may be, for example, a language model (e.g., a large-scale language model) having a Transformer structure.

[0024] As shown on the left side of FIG. 3, the model 140 has a first layer 301 , an intermediate layer A 300 , and a third layer 302 .

[0025] The first layer 301 is a layer that includes an input layer. The first layer 301 may include an intermediate layer other than the intermediate layer A300. The third layer 302 is a layer that includes an output layer. The third layer 302 may include an intermediate layer other than the intermediate layer A300.

[0026] In this embodiment, there is a model (referred to as the original model) that is the basis of model 140, and an anti-teacher mechanism is added to the intermediate layer A of the original model to form intermediate layer A300. That is, in this embodiment, an anti-teacher mechanism is added to a trained neural network model (a model corresponding to a predetermined process), and the anti-teacher mechanism controls the output, thereby making it possible to control the output so as to suppress undesirable output using the computational resources of one model (i.e., at a lower cost than conventional techniques).

[0027] As shown on the right side of Fig. 3, the hidden layer A300 includes an original model hidden layer A310 and an anti-teacher mechanism 320. As shown on the right side of Fig. 3, the anti-teacher mechanism 320 is connected in parallel to the original model hidden layer A310. Because it is added in parallel, the anti-teacher mechanism 320 can be added without any restrictions on the structure of the original model.

[0028] As described above, the anti-supervisory mechanism 320 can be added without any restrictions on the structure of the original model, and the intermediate layer A is not limited to a specific layer, but the intermediate layer A may be, for example, a Feed Forward (fully connected) layer or an Attention layer of a Transformer.

[0029] As shown on the right side of Figure 3, the output of the intermediate layer 310 of the original model is added to the output of the anti-teaching mechanism 320, and the value (feature) obtained by this addition is propagated to the next layer (the third layer in the case of Figure 3).

[0030] The anti-teaching mechanism 320 may be added to each of all hidden layers in the original model, or may be added to one or more hidden layers that are part of all hidden layers in the original model.

[0031] The anti-teaching mechanism 320 includes a vector conversion mechanism 321, a router 322, and a mode control mechanism 323. The processing contents of the anti-teaching mechanism 320 and its components will be explained below. In the explanation of each component, the input and output will be clearly stated, followed by an explanation of the processing contents.

[0032] <Anti-teacher mechanism 320> First, we will explain the anti-teacher mechanism 320. The input to the anti-teacher mechanism 320 is the input to the hidden layer A300, and the output from the anti-teacher mechanism 320 is a vector of the same dimension as the output from the hidden layer A310 of the original model.

[0033] The anti-teaching mechanism 320 inputs the input to the hidden layer A300 to a router 322 and a vector conversion mechanism 321, which will be described later.

[0034] Furthermore, the anti-teaching mechanism 320 multiplies the output vector from the vector conversion mechanism 321 by the output from the router 322 and the output from the mode control mechanism 323, and outputs the result.

[0035] The anti-teaching mechanism 320 is a trained model that converts input to output. The training method for the model will be described later.

[0036] <Vector Conversion Mechanism 321> Next, we will explain the vector conversion mechanism 321 in the anti-teacher mechanism 320. The input to the vector conversion mechanism 321 is the input to the hidden layer A300, and the output from the vector conversion mechanism 321 is a vector of the same dimension as the output from the hidden layer A310 of the original model.

[0037] The vector conversion mechanism 321 converts the input to the intermediate layer A300 and outputs a vector of the same dimension as the output of the intermediate layer A310 of the original model. The vector conversion mechanism 321 is a trained model, and this model as the vector conversion mechanism 321 converts the input to the output.

[0038] The structure of the vector conversion mechanism 321 as a neural network is not limited to a specific structure, but the vector conversion mechanism 321 can be realized, for example, by a fully connected layer in which the number of dimensions of the input and output is the same as the number of dimensions of the input to the intermediate layer A300.

[0039] In addition, the vector conversion mechanism 321 has the role of outputting a vector for changing the output of the entire original model to an undesired output, or the role of outputting a vector for changing the output of the entire original model to a desirable output.

[0040] <Router 322> Next, we will explain the router 322 in the anti-teacher mechanism 320. The input to the router 322 is the input to the hidden layer A 300. The output from the router 322 is a vector having a value between 0 and 1 for each element.

[0041] The router 322 is a trained model, and the model as the router 322 converts input to output. It is assumed that the router 322 outputs a vector indicating to what extent the anti-teacher mechanism 320 should be taken into consideration in the input to the intermediate layer A300.

[0042] The structure of router 322 as a neural network is not limited to a specific structure, but router 322 can be realized, for example, using a fully connected layer whose input dimensionality is the same as that of the input to intermediate layer A300 and whose output dimensionality is two, and a structure in which a softmax function is applied to the output of this fully connected layer to output a zeroth-dimensional vector.

[0043] <Mode Control Mechanism 323> The mode control mechanism 323 outputs −1. More specifically, during inference, the mode control mechanism 323 always outputs −1 to make the model 140 output data similar to the correct answer data in the learning phase.

[0044] Furthermore, if the parameters are adjusted during learning (the second learning described below) so that when the output of mode control mechanism 323 is 1, an undesirable output is produced for model 140 as a whole, and when the output of mode control mechanism 323 is -1, a desirable output is produced for model 140 as a whole, then, as described above, mode control mechanism 323 will always output -1 during inference.

[0045] This is not limited to this, and if the parameters are adjusted so that during learning (the second learning described below), when the output of mode control mechanism 323 is -1, an undesirable output is produced for the entire model 140, and when the output of mode control mechanism 323 is 1, a desirable output is produced for the entire model 140, then during inference, mode control mechanism 323 will always output 1.

[0046] (Configuration example of learning device 200) Fig. 4 shows a configuration example of the learning device 200 in this embodiment. As shown in Fig. 4, the learning device 200 includes an input unit 210, an inference unit 220, an output unit 230, and a parameter update unit 240. Fig. 4 also shows a trained model 260 to be output. Note that the "inference unit 220 + parameter update unit 240" may be referred to as a learning unit.

[0047] The learning device 200 is a device that performs (supervised) learning of a neural network model so as to receive predetermined data (information) as input and output the predetermined data (information). The learning device 200 is also a device that performs (supervised) learning of a neural network model so as to receive predetermined data (information) as input and output predetermined correct answer data (which may also be referred to as correct output data).

[0048] The operation of the learning device 200 is as follows: The following operation is applicable to any model in supervised learning assumed in this embodiment.

[0049] Here, it is assumed that the inference unit 220 already holds a model to be learned. Learning data for supervised learning is input to the input unit 210. The learning data is a pair of input data to the model and correct output data from the model for the input data.

[0050] Input data for the model is input from the input unit 210 to the inference unit 220. Also, correct answer output data is input from the input unit 210 to the parameter update unit 240.

[0051] The inference unit 220 obtains output data from the model for the input data, and passes the output data to the parameter update unit 240 .

[0052] The parameter update unit 240 uses a loss function to update the model parameters so as to reduce the error between the output data obtained from the model and the correct output data. As a parameter update method, a general machine learning method (e.g., backpropagation) can be used.

[0053] By repeating the above process, when learning is completed, the inference unit 220 outputs the learned model 260 via the output unit 230.

[0054] The learning process in this embodiment will be described below in accordance with the procedure of the flowchart in FIG.

[0055] <S201> In S201, the learning device 200 performs a first learning (first learning) on ​​the model using the learning method described above. The first learning may also be called pre-learning. The model to be learned here is the original model. That is, in the example of FIG. 3, the model has "a first layer 301, an intermediate layer A 310 of the original model, and a third layer 302."

[0056] The trained model trained in S201 is called trained model 1. In this embodiment, trained model 1 may be provided in advance. If trained model 1 is provided in advance, training in S201 is not performed. A model not trained in S201 may be called "model 1."

[0057] <S202> In S202, an anti-teacher mechanism (neural network model), which is a mechanism that plays the role of an anti-expert, is added to Trained Model 1. The "Trained Model 1 + anti-teacher mechanism" obtained in this way is called Trained Model 1'.

[0058] The method for generating the trained model 1' is not limited to a specific method, but for example, the trained model 1' may be generated using a learning device 400 equipped with an adding unit 250, as shown in Figure 6.

[0059] 6 is the same as learning device 200 except for the inclusion of adding unit 250. In other words, the model learning method itself is the same between learning device 400 and learning device 200.

[0060] 6, when adding an anti-teacher mechanism, model 1 (trained model 1) (270) is input to the input unit 210. When a name ends with a number, parentheses are used when describing the reference numeral in the drawing.

[0061] Since the input to the learning device 400 when adding is not limited to the trained model 1 that underwent the first learning in S201, it is written as "Model 1 (Trained Model 1)" in Figure 6.

[0062] The adding unit 250 generates a trained model 1' (280) by adding an anti-teacher mechanism 320 to the model (trained model 1) (270), and outputs the trained model 1' (280) to the inference unit 220. The inference unit 220 holds the trained model 1'.

[0063] As described above, the model input to the input unit 210 may be a model trained using the learning device 200 shown in FIG. 5 (trained model 1), or may be a pre-given model (model 1).

[0064] Furthermore, trained model 1' may also be generated by performing additional processing on model 1 (trained model 1) as described above, or trained model 1' may be provided in advance. The second learning described below uses a learning device 400 having an adding unit 250 shown in FIG. 6 , but when a pre-given trained model 1' is used as the learning target, learning may also be performed using the learning device 200 shown in FIG. 5 . As described above, the learning method itself is the same for learning device 200 and learning device 400.

[0065] <S203> In S203, the learning device 400 performs second learning by learning a model (trained model 1'). The second learning may also be called additional learning or fine-tuning. Since model 1 (trained model 1') is a model of model 140 shown in FIG. 3 in a state before the second learning, the trained model 1' will be referred to as model 140 here.

[0066] <S203: Overall Processing of Second Learning> First, the overall processing of the second learning will be described. Learning data for supervised learning is input to the input unit 210. The learning data is a pair of input data to the model 140 and correct output data from the model 140 for the input data.

[0067] Input data for the model 140 is input from the input unit 210 to the inference unit 220. Also, correct answer output data is input from the input unit 210 to the parameter update unit 240.

[0068] The inference unit 220 obtains output data from the model 140 in response to the input data, and passes the output data to the parameter update unit 240. The parameter update unit 240 updates the model parameters so as to reduce the error between the output data obtained from the model 140 and the correct output data. A typical machine learning method (e.g., backpropagation) can be used as a parameter update method. By repeating the above process, when learning is completed, the inference unit 220 outputs the trained model 140 (referred to as trained model 2) via the output unit 230.

[0069] The learning data in the first learning and the learning data in the second learning may be the same or different.

[0070] Furthermore, in the second learning, as will be described later, unlikelihood loss and cross entropy loss are used as loss functions to learn the entire model 140. Note that learning with unlikelihood loss may be followed by learning with cross entropy loss, or learning may be performed simultaneously. Unlikelihood loss may be called unlikelihood loss, and cross entropy loss may be called cross entropy loss.

[0071] In addition, in the second learning, the parameters of the original model may be fixed and the parameters of the added anti-teacher mechanism may be updated, which enables efficient learning and reduces the cost of the second learning (additional learning).

[0072] <S203: Operation of Model 140> The operation of model 140 during learning is the same as the operation of model 140 during inference, except for the operation of mode control mechanism 323. The operation of mode control mechanism 323 during learning will be described below.

[0073] During learning, the mode control mechanism 323 outputs 1 or −1. That is, the mode control mechanism 323 outputs 1 when learning the model 140 so that the model 140 can output an answer different from the correct answer data, and outputs −1 when learning the model 140 so that the model 140 can output the correct answer data. Specifically, for example, the mode control mechanism 323 outputs 1 and −1 for all samples included in the learning data, so that each case can be learned.

[0074] Note that 1 and −1 may be reversed. That is, the mode control mechanism 323 may output −1 when training the model 140 so that the model 140 can output an answer different from the correct answer data, and may output 1 when training the model 140 so that the model 140 can output the correct answer data.

[0075] Furthermore, if it is possible to distinguish between "producing an undesired output" and "producing a desired output", numerical values ​​other than 1 and -1 (for example, 2, -2) may be used as the output of the mode control mechanism 323.

[0076] That is, the learning unit performs learning to make the model 140 produce an undesired output when the output from the mode control mechanism 323 in the anti-teacher mechanism 320 is a first value, and performs learning to make the model 140 produce a desirable output when the output from the mode control mechanism 323 is a second value. The learning unit also switches the loss function based on the output from the mode control mechanism 323. Examples of loss functions include the aforementioned unlikelihood loss and cross entropy loss, but are not limited to these.

[0077] <S203: Operation of Parameter Update Unit 240> Next, the operation of the parameter update unit 240 in the second learning will be described. The parameter update unit 240 receives output data from the inference unit 220 (model 140) and correct answer output data corresponding to the input data to the inference unit 220. The parameter update unit 240 outputs the updated parameters of the model 140. As a result, the parameters of the model 140 in the inference unit 220 are updated with the updated parameters. The processing of the parameter update unit 240 in the second learning will be described in more detail below.

[0078] When the output from the mode control mechanism 323 is 1, the parameter updater 240 adjusts (updates) the parameters so that the output from the model 140 becomes an undesired output. When the output from the mode control mechanism 323 is −1, the parameter updater 240 adjusts (updates) the parameters so that the output from the model 140 becomes a desirable output.

[0079] More specifically, the parameter update unit 240 optimizes the parameters using a loss function that receives as input the output data from the inference unit 220 and the corresponding correct output data. An example of specific processing is as follows.

[0080] The parameter update unit 240 performs learning for all samples included in the training data in cases where the output of the mode control mechanism 323 is both 1 and −1. When the output from the mode control mechanism 323 is 1, the parameter update unit 240 uses the unlikelihood loss as the loss function, and when the output of the mode control mechanism 323 is −1, the parameter update unit 240 uses the cross entropy loss as the loss function. Note that, based on an instruction from the parameter update unit 240, the mode control mechanism 323 may determine whether to output 1 or −1.

[0081] That is, the parameter update unit 240 adjusts the parameters using the unlikelihood loss as a loss function when the output of the mode control mechanism 323 is 1, and adjusts the parameters using the cross entropy loss as a loss function when the output of the mode control mechanism 323 is −1.

[0082] Unlikelihood loss is a loss function that allows the model to output an answer different from the correct answer data, and cross entropy loss is a loss function that allows the model to output the correct answer data.

[0083] As a specific loss function for the unlikelihood loss, for example, the loss function disclosed in "Neural Text Generation with Unlikelihood Training (https: / / arxiv.org / abs / 1908.04319)" can be used. However, the specific function used for the unlikelihood loss is not limited to a specific function. Any function that increases the loss when the model output is correct and decreases the loss when the model output is incorrect can be used.

[0084] The parameters to be learned (updated) by the parameter update unit 240 may be parameters of the entire model 140, or may be parameters of the anti-teacher mechanism 320 in the model 140 (the parameters of other parts of the model 140 are fixed).

[0085] More specifically, the minimum necessary parameters to be updated during learning may be the parameters of the vector conversion mechanism 321 and the router 322 in the anti-teacher mechanism 320. The mode control mechanism 323 may not have learnable parameters.

[0086] <S203: Specific example of second learning> Assume that the model 140 is a large-scale language model, and it is desired to improve the truthfulness of the model 140. Using a data set consisting of a pair of a question and its answer, the input of the model 140 is the question, and the output is the answer.

[0087] For example, if the input question is "What is the capital of Canada?" and the correct answer is "Ottawa," the input question is provided as an input to model 140, and the correct answer is provided as correct data for the output from model 140. The output of mode control mechanism 323 is set to -1, and parameters are optimized using cross entropy loss. Also, the output from mode control mechanism 323 is set to 1, and parameters are optimized using unlikelihood loss.

[0088] As described above, in the second learning, the parameters are adjusted so that when the output of the mode control mechanism 323 is 1, an undesirable output is produced for the entire model, and when the output of the mode control mechanism 323 is -1, a desirable output is produced for the entire model, thereby improving the accuracy of verisimilitude compared to the original model. The reason for the improved accuracy is that the model can (implicitly) recognize undesirable outputs.

[0089] (Hardware Configuration Example) Any of the devices described in this embodiment (inference device 100, learning device 200, learning device 400, information processing device) can be realized, for example, by causing a computer to execute a program. This computer may be a physical computer or a virtual machine on the cloud.

[0090] That is, the device can be realized by executing a program corresponding to the processing performed by the device using hardware resources such as a CPU and memory built into a computer. The program can be recorded on a computer-readable recording medium (such as a portable memory) and stored or distributed. The program can also be provided via a network such as the Internet or email.

[0091] Fig. 7 is a diagram showing an example of the hardware configuration of the computer. The computer in Fig. 7 includes a drive device 1000, an auxiliary storage device 1002, a memory device 1003, a CPU 1004, an interface device 1005, a display device 1006, an input device 1007, an output device 1008, and the like, all of which are interconnected via a bus B. The computer may further include a GPU.

[0092] The program that realizes the processing on the computer is provided by a recording medium 1001, such as a CD-ROM or a memory card. When the recording medium 1001 storing the program is set in the drive device 1000, the program is installed from the recording medium 1001 to the auxiliary storage device 1002 via the drive device 1000. However, the program does not necessarily have to be installed from the recording medium 1001, but may be downloaded from another computer via a network. The auxiliary storage device 1002 stores the installed program as well as necessary files, data, etc.

[0093] The memory device 1003 reads and stores a program from the auxiliary storage device 1002 when an instruction to start the program is received. The CPU 1004 realizes functions related to the device in accordance with the program stored in the memory device 1003. The interface device 1005 is used as an interface for connecting to a network, etc. The display device 1006 displays a GUI (Graphical User Interface) or the like according to the program. The input device 1007 is composed of a keyboard, mouse, buttons, a touch panel, etc., and is used to input various operation instructions. The output device 1008 outputs the results of calculations.

[0094] (Effects of the embodiment) As described above, the technology described in this embodiment provides a technology for suppressing undesirable outputs from a trained neural network model at a lower cost than conventional technology.

[0095] More specifically, the technology described in this embodiment uses a single model in which an anti-teaching mechanism is added to an original model, rather than using two models as in the conventional technology (Non-Patent Document 1), thereby reducing computational resources and enabling control to suppress undesirable outputs without extending the inference time.

[0096] An experiment was conducted to verify the effectiveness of the technology described in this embodiment, and the experimental results are shown in FIG. 8 . In this experiment, the QA domain (10,000 samples) of HaluEval (TruthfulQA: Measuring How Models Mimic Human Falsehoods (https: / / arxiv.org / abs / 2109.07958)) was used as the training dataset, and TruthfulQA (HaluEval: A Large-Scale Hallucination Evaluation Benchmark for Large Language Models (https: / / arxiv.org / abs / 2305.11747)) was used as the evaluation dataset. Furthermore, the 7B model of Llama2 (https: / / llama.meta.com / llama2 / ) was used as the model. Furthermore, the intermediate layer A was a feed-forward layer.

[0097] MC1, MC2, and MC3 were used as evaluation indices. A higher score indicates a higher degree of truthfulness. MC stands for Multiple-Choice.

[0098] As a result of the experiment, the technology according to this embodiment (proposed method) was able to produce results similar to those of the original model using the computational resources of one model. In particular, in the case shown in Figure 8, it can be seen that the veracity is improved compared to the original model.

[0099] The following additional notes are provided regarding the above-described embodiments.

[0100] <Additional Notes> (Additional Item 1) An information processing device that trains a neural network model, comprising: a memory; and at least one processor connected to the memory, wherein the processor trains the model to cause the model to produce an undesired output, and to cause the model to produce a desired output. (Additional Item 2) The information processing device according to Additional Item 1, wherein the model has an intermediate layer including an anti-teacher mechanism. (Additional Item 3) The information processing device according to Additional Item 1, wherein the model has a mode control mechanism that controls whether the model trains to cause the model to produce an undesired output, or to cause the model to produce a desired output. (Additional Item 4) The information processing device according to Additional Item 2, wherein the model has a structure in which the anti-teacher mechanism is arranged in parallel with an intermediate layer of a model corresponding to a predetermined process. (Additional Item 5) An information processing device comprising: a memory; and at least one processor connected to the memory, wherein the processor performs inference using a neural network model, wherein the model has a structure in which an anti-teacher mechanism is arranged in parallel with an intermediate layer of a model corresponding to the predetermined process. (Supplementary Item 6) A learning method executed by an information processing device that learns a neural network model, the learning method comprising a learning step of training the model to make the model produce an undesired output and training the model to make the model produce a desirable output. (Supplementary Item 7) A non-transitory storage medium that stores a program for causing a computer to function as the information processing device according to any one of Supplementary Items 1 to 5.

[0101] Although the present embodiment has been described above, the present invention is not limited to such a specific embodiment, and various modifications and changes are possible within the scope of the gist of the present invention described in the claims.

[0102] 100 Inference device 110 Input unit 120 Inference unit 130 Output unit 140 Model 200, 400 Learning device 210 Input unit 220 Inference unit 230 Output unit 240 Parameter update unit 250 Addition unit 260 Trained model 270 Model 1 (Trained model 1) 280 Trained model 1' 300 Hidden layer A 301 First layer 302 Third layer 310 Hidden layer A of original model 320 Anti-teacher mechanism 321 Vector conversion mechanism 322 Router 323 Mode control mechanism 1000 Drive device 1001 Recording medium 1002 Auxiliary storage device 1003 Memory device 1004 CPU 1005 Interface device 1006 Display device 1007 Input device 1008 Output device

Claims

1. An information processing device that trains a neural network model, comprising a learning unit that trains the model to make it produce undesirable output and to make it produce desirable output.

2. The information processing device according to claim 1, wherein the model has an intermediate layer including an anti-teaching mechanism.

3. An information processing device according to claim 1, wherein the model has a mode control mechanism for controlling whether the model is trained to produce an undesired output or trained to produce a desired output.

4. The information processing device according to claim 2, wherein the model has a structure in which the anti-teaching mechanism is arranged in parallel with an intermediate layer of a model corresponding to a predetermined process.

5. An information processing device comprising an inference unit that performs a predetermined process using a neural network model, the model having a structure in which an anti-teacher mechanism is arranged in parallel to an intermediate layer of the model corresponding to the predetermined process.

6. A learning method executed by an information processing device that trains a neural network model, comprising a learning step of training the model to make it produce an undesirable output and training the model to make it produce a desirable output.

7. An inference method executed by an information processing device that performs a specified processing using a neural network model, wherein the model has a structure in which an anti-teacher mechanism is arranged in parallel to an intermediate layer of the model corresponding to the specified processing, and the specified processing is performed using the model.

8. A program for causing a computer to function as each unit in the information processing device according to any one of claims 1 to 5.

Images