Information processing device and information processing method

Abstract

An information processing device comprises a creation unit, an extraction unit, and a prediction unit. The creation unit creates a set having, as an element, a combination that satisfies a condition based on hierarchy information indicating the hierarchy of concepts among combinations of the concepts and the hierarchy information. The extraction unit extracts, from the set, a concept for which the magnitude of the degree of similarity to input data satisfies a condition. The prediction unit predicts output data corresponding to the input data on the basis of the concept extracted by the extraction unit.

Description

Information processing device and information processing method

[0001] The present invention relates to an information processing apparatus and an information processing method.

[0002] Conventionally, concept bottleneck models (CBMs) are known, which predict concepts, which are constituent elements of data, from the intermediate output of a neural network, and output the final predicted label from the predicted concepts (see, for example, Non-Patent Document 1).

[0003] Figure 10 illustrates a conventional concept bottleneck. As shown in Figure 10, in a concept bottleneck, the DNN (Deep Neural Network) outputs the label "Lynx" for the input image data. At this time, a predetermined score for each concept is obtained from the DNN's intermediate output. For example, "feline" and "large arms" are concepts. With a concept bottleneck, the concepts that respond to the final label prediction become explicit, making it possible to interpret the output.

[0004] Koh, Pang Wei, et al. "Concept bottleneck models." ICML 2020Ramaswamy, Vikram V., et al. "Overlooked factors in concept-based explanations: Dataset choice, concept learnability, and human capability." CVPR 2023

[0005] However, conventional technologies have the problem of not being able to efficiently address concept bottlenecks.

[0006] For example, the concept bottleneck described in Non-Patent Document 1 requires training for each model and target dataset. For instance, training a concept bottleneck requires constructing a set of concepts tailored to the target dataset.

[0007] Furthermore, it is known that as the number of concepts increases, learning both concepts and task labels becomes difficult, making it challenging to realize general-purpose CBMs through training on large datasets (see, for example, Non-Patent Document 2).

[0008] This invention has been made in view of the above, and aims to efficiently address the concept bottleneck.

[0009] To solve the above-mentioned problems and achieve the objective, the information processing device according to the present invention is characterized by comprising: a creation unit that creates a set of elements which are combinations of a concept and hierarchical information indicating the hierarchy of the concept, and which satisfy the conditions based on the hierarchical information; an extraction unit that extracts from the set concepts whose similarity to input data satisfies the conditions; and a prediction unit that predicts output data corresponding to the input data based on the concepts extracted by the extraction unit.

[0010] According to the present invention, it becomes possible to efficiently address concept bottlenecks.

[0011] Figure 1 is a diagram showing an example configuration of an information processing device according to the first embodiment. Figure 2 is a diagram illustrating the overview of the model. Figure 3 is a diagram showing the algorithm for concept hierarchization. Figure 4 is a diagram showing the algorithm for concept filtering. Figure 5 is a flowchart showing the flow of the learning process. Figure 6 is a flowchart showing the flow of the inference process. Figure 7 is a flowchart showing the flow of the concept hierarchization process. Figure 8 is a flowchart showing the flow of the concept filtering process. Figure 9 is a diagram showing an example of a computer that executes an information processing program. Figure 10 is a diagram illustrating a conventional concept bottleneck.

[0012] Hereinafter, one embodiment of the present invention will be described in detail with reference to the drawings. However, the present invention is not limited to this embodiment. Furthermore, in the drawings, the same parts are denoted by the same reference numerals.

[0013] [First Embodiment] The configuration of the information processing device according to the first embodiment will be described using Figure 1. Figure 1 is a diagram showing an example of the configuration of the information processing device according to the first embodiment.

[0014] The information processing device 10 is a device for training a machine learning model and performing inference using the machine learning model. In this embodiment, the machine learning model (hereinafter simply referred to as the model) performs the task of classifying images. However, the task performed by the model may be a task other than image classification, such as time series data prediction or speech recognition. Also, in this embodiment, the execution of a task using the model is referred to as prediction, regardless of the type of task.

[0015] The information processing device 10 in this embodiment can not only perform tasks but also obtain concepts corresponding to the input data. However, in this embodiment, it is not necessary to update the neural network parameters to match the target task.

[0016] As shown in Figure 1, the information processing device 10 includes a communication unit 11, an input unit 12, an output unit 13, a storage unit 14, and a control unit 15.

[0017] The communication unit 11 performs data communication with other devices via a network. For example, the communication unit 11 is a NIC (Network Interface Card). The input unit 12 is an interface connected to input devices such as a keyboard and mouse. The output unit 13 is an interface connected to output devices such as a display.

[0018] The storage unit 14 is a storage device such as an HDD (Hard Disk Drive), SSD (Solid State Drive), or optical disc. Alternatively, the storage unit 14 may be a rewritable semiconductor memory such as RAM (Random Access Memory), flash memory, or NVSRAM (Non-Volatile Static Random Access Memory). The storage unit 14 stores the OS (Operating System) and various programs executed by the information processing device 10.

[0019] The memory unit 14 stores model information 141 and concept information 142. Model information 141 consists of model parameters. For example, model information 141 includes the weights and biases of a neural network. Concept information 142 is a set of concepts, which will be described later.

[0020] The control unit 15 controls the entire information processing device 10. The control unit 15 is, for example, an electronic circuit such as a CPU (Central Processing Unit), MPU (Micro Processing Unit), or GPU (Graphics Processing Unit), or an integrated circuit such as an ASIC (Application Specific Integrated Circuit) or FPGA (Field Programmable Gate Array).

[0021] Furthermore, the control unit 15 has an internal memory for storing programs and control data that define various processing procedures, and executes each process using the internal memory. In addition, the control unit 15 functions as various processing units when various programs are run. For example, the control unit 15 has a creation unit 151, a conversion unit 152, an extraction unit 153, a prediction unit 154, an update unit 155, and a provision unit 156.

[0022] Here, the configuration of the model and the processing of each part of the control unit 15 will be explained. Figure 2 is a diagram illustrating the overview of the model. As shown in Figure 2, the model 20 has a Vision Encoder 21, a General Concept Bank 22, a Regressor 23, and a Classifier 24. The model 20 is constructed based on the model information 141. Note that the model 20 does not require training according to the target task and can extract concepts in zero shots, so it may also be called a Zero-shot Concept Bottleneck Model (Z-CBM).

[0023] Vision Encoder 21 encodes the input data x. The input data x is, for example, image data. For example, Vision Encoder 21 is CLIP (see Reference 1), a pre-trained visual language infrastructure model.

[0024] Reference 1: Radford, Alec, et al. "Learning transferable visual models from natural language supervision." International conference on machine learning. PMLR, 2021.

[0025] General Concept Bank 22 is a predefined set of concepts. For example, General Concept Bank 22 may be an existing set of concepts defined by a human (Reference 2), or it may be a set of concepts automatically generated by an LLM (Large Language Models) (Reference 3).

[0026] Reference 2: Speer, Robyn, Joshua Chin, and Catherine Havasi. "Conceptnet 5.5: An open multilingual graph of general knowledge. singh 2002 (2016)." AAAI 2017.

[0027] Reference 3: Brown, Tom B. "Language models are few-shot learners." NeurIPS 2020.

[0028] Regressor 23 performs regression weighting on candidate concepts (Top-K Concepts) extracted from General Concept Bank 22. Classifier 24 outputs labels corresponding to the input data x as the final prediction result. For example, "Lynx" is a label. Note that Classifier 24 may be replaced with a model appropriate for the task.

[0029] Returning to Figure 1, the creation unit 151 creates a set of concepts (General Concept Bank 22). The transformation unit 152 encodes the input data using an encoder (Vision Encoder 21). In other words, the transformation unit 152 converts the input data into a vector of a specified format.

[0030] The extraction unit 153 extracts concepts from the set of concepts based on the encoding results. The prediction unit 154 uses the Classifier 24 to predict output data (e.g., labels) based on the extracted concepts. In other words, the prediction unit 154 predicts output data corresponding to the input data based on the concepts extracted by the extraction unit 153.

[0031] The update unit 155 updates the parameters of the Classifier 24 based on the predicted output data. For example, the update unit 155 updates the parameters of the neural network using backpropagation.

[0032] Here, in order to obtain a concept to explain arbitrary input data, it is necessary to use a large set of concepts containing a variety of vocabulary. For example, a concept can be a word, a phrase, or something similar.

[0033] On the other hand, the concepts included in a concept collection are not always the level of detail that users desire. For example, when describing an "apple," sometimes highly abstract information such as "red circle" is desirable, while other times a specific variety name such as "Fuji" is preferable.

[0034] Therefore, the creation unit 151 controls the granularity by mapping the elements within the large-scale concept set to an existing hierarchical structure and specifying the relative granularity to the output (e.g., labels) in the task. The existing hierarchical structure may be based on, for example, WordNet (Reference 4). For example, the granularity (hierarchical information) is hypernym, hyponym, and co-hyponym.

[0035] Reference 4: George A. Miller (1995). WordNet: A Lexical Database for English. Communications of the ACM Vol. 38, No. 11: 39-41.

[0036] Therefore, the creation unit 151 performs hierarchical organization on a large set of concepts given in advance. Furthermore, the creation unit 151 filters the set of concepts based on a specified granularity. The General Concept Bank 22 is the set of concepts obtained through the hierarchical organization and filtering performed by the creation unit 151.

[0037] Concept hierarchy will be explained using Figure 3. Figure 3 is a diagram showing the algorithm for concept hierarchy. ref This is a set of words for reference. ref Each word in C is associated with hierarchical information. C is a set of concepts. The elements of C are combinations of a word c and hierarchical information. However, in the initial state, at least some of the elements of C may be words only and not have hierarchical information assigned to them.

[0038] The creation unit 151 performs the following processing on each of the words c contained in C (second and seventh lines in Figure 3). First, the creation unit 151 processes W ref The nearest neighboring word (reference word) to word c is obtained (line 4). The creation unit 151 combines the obtained reference word's hierarchical information w* with word c and stores it in the concept set Hc (lines 5 and 6).

[0039] In this way, the creation unit 151 retrieves the word with the greatest similarity to the concept, which is a word, from the reference set, which is a set of words associated with hierarchical information, and adds the hierarchical information associated with the retrieved word to the set in combination with the concept.

[0040] The filtering of the concept set will be explained using Figure 4. Figure 4 is a diagram showing the concept filtering algorithm. The upper limit of the hierarchical information h is set by the user. upper , and lower limit h loweris specified.

[0041] Creation unit 151 performs the following processing for each element of the hierarchically ordered concept set according to the algorithm of FIG. 3 (second and sixth lines of FIG. 4). Creation unit 151 uses the hierarchical information h of the elements of the concept set c If it is greater than the lower limit value and less than the upper limit value, the element is added to the filtered concept set C' (third and fourth lines).

[0042] The hierarchical information may be represented by a numerical value. For example, the numerical value of the hierarchical information becomes smaller as the hierarchical level is higher and larger as the hierarchical level is lower.

[0043] In this way, creation unit 151 creates a set having as elements combinations (for example, combinations within the range specified by the hierarchical information in filtering) that satisfy the conditions based on the hierarchical information among the combinations of concepts and hierarchical information indicating the hierarchy of the concepts.

[0044] By creating such a hierarchically ordered set, there is no need to prepare a learning dataset, and information processing apparatus 10 can directly perform inference based on the input data. Also, by concept hierarchical ordering and filtering, information processing apparatus 10 can control the granularity of the concepts.

[0045] Extraction unit 153 performs a vector search on the concept set based on the vector converted by conversion unit 152, that is, the vector output from the encoder, and extracts candidates for concepts. Extraction unit 153 performs the search using a fast and scalable search algorithm such as FAISS (Reference 5), for example.

[0046] Reference 5: Douze, Matthijs, et al. "The faiss library." arXiv preprint arXiv:2401.08281 (2024).

[0047] For example, the extraction unit 153 uses the vector output from the Vision Encoder 21 as a Query and searches the concept set for a predetermined number of concepts (e.g., K concepts) in descending order of similarity to the Query (Top-K Concepts). The elements of the concept set may be converted into vectors. The concepts obtained through the search are called the concept candidate set.

[0048] In this way, the extraction unit 153 extracts concepts from the set that satisfy the condition regarding the magnitude of similarity with the input data. By performing this concept search, the information processing device 10 can handle concept sets of any size.

[0049] Here, the set of candidate concepts retrieved based on similarity may be noisy and contain overlapping meanings. In that case, the inference accuracy in the task decreases. Therefore, the extraction unit 153 uses concept regression (Regressor 23) to assign weights to each of the concepts (vectors) included in the set of candidate concepts.

[0050] The extraction unit 153 calculates the weights W by sparse linear regression (e.g., Lasso) as shown in equations (1) and (2), and then calculates the weighted sum ^C using W. T Calculate W (where ^C is directly above C).

[0051]

[0052]

[0053] Furthermore, the prediction unit 154 uses the Classifier 24 to predict the labels as shown in equation (3).

[0054]

[0055] f in equation (1) V (x) is a vector of size d, which is the output of the Vision Encoder 21. K is the number of elements in the set of concept candidates ^C extracted by the extraction unit 153. W is a K × D dimension weight matrix. λ is a pre-set positive constant.

[0056] Equation (3) means that among the set Y of the original labels, the label y with the highest similarity to the weighted sum ^C T W is selected as the final output data. l() is Classifier 24. Thus, by using sparse linear regression, it becomes difficult to select concepts similar to each other during label prediction.

[0057] Thus, the extraction unit 153 obtains the respective weights of the vectors representing the extracted concepts by regression. Also, the prediction unit 154 predicts output data based on the weighted sum of the weights and the vectors. The extraction unit 153 can achieve the same inference accuracy as the original black box DNN by such concept regression.

[0058] [Processing of the First Embodiment] Using FIG. 5, the flow of the learning process by the information processing apparatus 10 will be described. FIG. 5 is a flowchart showing the flow of the learning process.

[0059] As shown in FIG. 5, first, the conversion unit 152 encodes the input data using an encoder (step S101). Subsequently, the extraction unit 153 searches a predefined set of concepts and extracts concept candidates based on the result of the encoding (step S102). That is, the extraction unit 153 searches for vectors with high similarity to the vector (Query) obtained as the result of the encoding.

[0060] The prediction unit 154 predicts the label corresponding to the input data by concept regression based on the extracted concept candidates (step S103).

[0061] The update unit 155 updates the parameters of the encoder and the concept regression based on the difference between the predicted label and the correct label (step S104). When the learning is completed (step S105: Yes) (for example, when S104 is executed a specified number of times), the information processing apparatus 10 ends the process. When the learning is not completed (step S105: No), the information processing apparatus 10 returns to step S101 and repeats the process.

[0062] The flow of inference processing by the information processing device 10 will be explained using Figure 6. Figure 6 is a flowchart showing the flow of inference processing.

[0063] As shown in Figure 6, first, the conversion unit 152 encodes the input data using an encoder (step S201). Next, the extraction unit 153 searches a predefined set of concepts and extracts concept candidates based on the encoding result (step S202). That is, the extraction unit 153 searches for vectors that have a high similarity to the vector (Query) obtained as a result of encoding.

[0064] The prediction unit 154 predicts labels corresponding to the input data by performing concept regression based on the extracted concept candidates (step S203).

[0065] Figure 7 illustrates the process flow of concept hierarchy. Figure 7 is a flowchart showing the process flow of concept hierarchy.

[0066] As shown in Figure 7, first, the creation unit 151 determines whether or not there are any elements in the concept set for which hierarchical information has not been registered (step S301). If there are no unregistered elements (step S301: No), the creation unit 151 terminates the process.

[0067] If there are unregistered elements (step S301: Yes), the creation unit 151 selects one element whose hierarchical information is not registered (step S302). Next, the creation unit 151 obtains a reference element, which is the nearest neighbor element to the selected element, from the concept set (step S303).

[0068] Here, the creation unit 151 acquires the hierarchical information registered in the reference element (step S304). Then, the creation unit 151 registers the hierarchical information in the selected element (step S305).

[0069] Figure 8 will be used to explain the processing flow of concept filtering. Figure 8 is a flowchart showing the processing flow of concept filtering.

[0070] As shown in Figure 8, first, the creation unit 151 determines whether or not there are any undetermined elements in the concept set (step S401). If there are no undetermined elements (step S401: No), the creation unit 151 terminates the process.

[0071] If there are undetermined elements (step S401: Yes), the creation unit 151 selects one of the undetermined elements (step S402).

[0072] Here, the creation unit 151 determines whether the hierarchy of the selected element is within a specified range (step S403). If it is within the specified range (step S403: Yes), the creation unit 151 adds the selected element to the filtered set (step S404). If it is not within the specified range (step S403: No), the creation unit 151 returns to step S401 without adding the selected element to the filtered set.

[0073] As explained above, the creation unit 151 creates a set of elements consisting of combinations of concepts and hierarchical information indicating the hierarchy of concepts, where the combinations satisfy the conditions based on the hierarchical information. The extraction unit 153 extracts from the set concepts whose similarity to the input data satisfies the conditions. The prediction unit 154 predicts output data corresponding to the input data based on the concepts extracted by the extraction unit 153. In this way, the information processing device 10 can efficiently implement concept bottlenecks by creating a hierarchical set of concepts.

[0074] [Program] In one embodiment, the information processing device 10 can be implemented by installing an information processing program that performs the above processing as packaged software or online software on a desired computer. For example, by having the computer execute the information processing program, the computer can be made to function as the information processing device 10.

[0075] Figure 9 shows an example of a computer that executes an information processing program. Computer 1000 has, for example, memory 1010 and CPU 1020. Computer 1000 also has a hard disk drive interface 1030, a disk drive interface 1040, a serial port interface 1050, a video adapter 1060, and a network interface 1070. These components are connected by a bus 1080.

[0076] Memory 1010 includes ROM (Read Only Memory) 1011 and RAM (Random Access Memory) 1012. ROM 1011 stores, for example, a boot program such as BIOS (Basic Input Output System). The hard disk drive interface 1030 is connected to the hard disk drive 1090. The disk drive interface 1040 is connected to the disk drive 1100. For example, a removable storage medium such as a magnetic disk or optical disk is inserted into the disk drive 1100. The serial port interface 1050 is connected to, for example, a mouse 1110 and a keyboard 1120. The video adapter 1060 is connected to, for example, a display 1130.

[0077] The hard disk drive 1090 stores, for example, the OS 1091, application programs 1092, program modules 1093, and program data 1094. That is, the programs that define each process of the information processing device 10 are implemented as program modules 1093 in which executable code for the computer is written. The program modules 1093 are stored, for example, in the hard disk drive 1090. For example, a program module 1093 for performing processes similar to the functional configuration of the information processing device 10 is stored in the hard disk drive 1090. Note that the hard disk drive 1090 may be replaced by an SSD.

[0078] Furthermore, the configuration data used in the processing of the above-described embodiment is stored as program data 1094 in, for example, memory 1010 or hard disk drive 1090. The CPU 1020 then reads the program module 1093 and program data 1094 stored in memory 1010 or hard disk drive 1090 into RAM 1012 as needed and executes the processing of the above-described embodiment.

[0079] Furthermore, the program module 1093 and program data 1094 are not limited to being stored in the hard disk drive 1090; for example, they may be stored in a removable storage medium and read by the CPU 1020 via a disk drive 1100 or the like. Alternatively, the program module 1093 and program data 1094 may be stored in another computer connected via a network (LAN (Local Area Network), WAN (Wide Area Network), etc.). The program module 1093 and program data 1094 may then be read by the CPU 1020 from the other computer via a network interface 1070.

[0080] Although embodiments applying the invention made by the present inventors have been described above, the present invention is not limited by the descriptions and drawings that constitute part of the disclosure of the present invention in this embodiment. That is, all other embodiments, examples, and operational techniques made by those skilled in the art based on this embodiment are included in the scope of the present invention.

[0081] 10 Information Processing Device 11 Communication Unit 12 Input Unit 13 Output Unit 14 Storage Unit 15 Control Unit 20 Model 21 Vision Encoder 22 General Concept Bank 23 Regressor 24 Classifier 141 Model Information 142 Concept Information 151 Creation Unit 152 Conversion Unit 153 Extraction Unit 154 Prediction Unit 155 Update Unit 156 Provision Unit

Claims

1. An information processing device comprising: a creation unit that creates a set of elements which are combinations of a concept and hierarchical information indicating the hierarchy of the concept, and which satisfy the conditions based on the hierarchical information; an extraction unit that extracts from the set concepts whose similarity to the input data satisfies the conditions; and a prediction unit that predicts output data corresponding to the input data based on the concepts extracted by the extraction unit.

2. The information processing apparatus according to claim 1, characterized in that the creation unit obtains the word with the greatest similarity to the concept, which is a word, from a reference set, which is a set of words associated with hierarchical information, and adds the hierarchical information associated with the obtained word to the set in combination with the concept.

3. The information processing apparatus according to claim 1, characterized in that the extraction unit obtains the weights of each of the extracted vectors representing the concepts by regression, and the prediction unit predicts the output data based on the weighted sum of the weights and the vectors.

4. An information processing method performed by an information processing device, comprising: a creation step of creating a set of elements which are combinations of a concept and hierarchical information indicating the hierarchy of the concept, and which satisfy conditions based on the hierarchical information; an extraction step of extracting from the set concepts whose similarity to input data satisfies the conditions; and a prediction step of predicting output data corresponding to the input data based on the concepts extracted in the extraction step.

Images