Assessment method for continuous learning of multi-source heterogeneous data based on context knowledge enhancement

Abstract

The invention discloses a multi-source heterogeneous data continuous learning evaluation method based on context knowledge enhancement, and relates to a continuous learning technology and a model performance evaluation method in machine learning. The method mainly aims at evaluating the ability of a depth model to continuously learn sequence tasks on multi-source heterogeneous data; the method is an evaluation method which is based on context knowledge enhancement and focuses on key new knowledge learning efficiency and precision. According to the method, a human forgetting curve is simulated, the defect that new knowledge and old knowledge are not distinguished in an existing evaluation mode is broken through, higher calculation weights are given to knowledge needed by the model for processing problems in the current context environment, and the continuous learning ability of the model on multi-source heterogeneous data is evaluated in a mode more fitting the reality.

Description

Technical Field

[0001] This invention relates to continuous learning techniques and model performance evaluation methods in machine learning. Background Technology

[0002] In real-world applications, models are often required to learn sequential tasks. For example, a home care robot needs to possess cooking skills in a kitchen setting and perhaps the ability to clean and tidy a bedroom. After training in a kitchen setting, we expect the robot to retain the kitchen-related skills acquired during training in a bedroom setting, thus accumulating new and old knowledge. Currently, although deep learning models based on backpropagation gradient updates perform well on tasks with relatively complete data and similar training and deployment scenarios, these deep models face severe catastrophic forgetting when learning new tasks sequentially. That is, after learning a new task, the model loses a significant amount of the previously learned ability to handle older tasks. Therefore, endowing deep models with continuous learning capabilities is a crucial step in applying them to real-world scenarios.

[0003] Deploying deep models in real-world scenarios requires not only addressing the catastrophic forgetting problem described above, but also the ability to learn from diverse, heterogeneous data sources. For example, deep models deployed in home care robots need to process not only visual image data, but also data from speech, radar, and posture. Therefore, endowing deep models with the ability to process diverse, heterogeneous data sources is a key requirement for their deployment in real-world scenarios. Summary of the Invention

[0004] This invention attempts to address the problem of catastrophic forgetting by evaluating the knowledge extraction capability of a model after continuous learning on multi-source heterogeneous data.

[0005] The technical solution adopted by this invention to solve the above-mentioned technical problems is: an evaluation method for continuous learning of multi-source heterogeneous data based on contextual knowledge reinforcement, the method comprising:

[0006] Step 1: Initialize a multi-source heterogeneous data processing network Its input data includes image and text modalities, and its structure includes an image data encoder. Text data encoder Modal interaction structure and classifier network Text data input text data encoder Image data input image data encoder Text data encoder and image data encoder Output-input modal interaction structure Modal interaction structure The multivariate heterogeneous data fusion feature is output, and the multivariate heterogeneous data fusion feature is input into the classifier, which outputs the category.

[0007] Step 2: Process the dataset Preprocessing is performed to divide it into training sets. and test set The training set is divided into multiple subsets. The test set is divided according to the same criteria as the training set, resulting in... After this division, each subset of the training set and the subset of the test set constitute a dataset for a task. Multiple tasks sequentially fed into a multi-source heterogeneous data processing network constitute a task sequence. ; Indicates the total number of tasks;

[0008] Step 3: In the task During training, data is used Training data in Training a multi-source heterogeneous data processing network, based on training data Data from each batch of samples The loss is calculated, and the parameters of the multi-source heterogeneous data processing network are optimized by backpropagating gradient updates.

[0009] Step 4: In the task After training is completed, test data from all previously learned tasks will be used. The above tests the multi-source heterogeneous data processing network trained so far. And calculate the relevant accuracy, and obtain ;

[0010] Step 5: Repeat steps 3 and 4 until the task sequence is complete. On-the-job training;

[0011] Step 6: Collect the accuracy rates calculated in Step 4 to obtain the accuracy matrix;

[0012] The average accuracy of the model across all tasks learned at that time is obtained by averaging the lower triangular portion of the accuracy matrix. This average accuracy is denoted as:

[0013] ;

[0014] Step 7: Based on the human forgetting curve ,in For the probability of recall, For the strength of memory traces, Set a knowledge weight matrix for time. ;

[0015] Step 8: Calculate the average incremental accuracy of contextual knowledge reinforcement:

[0016] .

[0017] Furthermore, in step 2, the training set is divided into multiple subsets based on task differences, domain differences, or category differences.

[0018] Furthermore, the accuracy matrix in step 6 is as follows: .

[0019] Furthermore, in step 7: ;

[0020] in, It is a natural constant. Indicates the task number.

[0021] The evaluation method for continuous learning of multi-source heterogeneous data based on contextual knowledge reinforcement proposed in this invention is inspired by the human forgetting law. By assigning higher weights to the contextual task currently being processed by the model and designing an exponential decay curve to weaken the weight of historical tasks in the calculation of evaluation indicators, an average incremental accuracy index for contextual knowledge reinforcement is proposed, realizing an evaluation method that focuses on new knowledge learning for multi-source heterogeneous data processing models. Attached Figure Description

[0022] Figure 1 This is a basic framework diagram of a multi-source heterogeneous data processing model. Detailed Implementation

[0023] This invention is implemented on the PyTorch deep learning framework and mainly includes the following steps:

[0024] Phase 1:

[0025] Step 1: Define a multi-source heterogeneous data processing network, where the image data encoder and text data encoder are initialized with pre-trained weights, and the remaining modal interaction structures and classifier networks are randomly initialized;

[0026] Step 2: Divide the dataset into dataset sequences corresponding to the task sequences, and accordingly divide it into training and test sets;

[0027] Step 3: Train the multi-source heterogeneous data processing network on the training data of the current task, and calculate the accuracy of the trained network on the test data of all the tasks it has learned.

[0028] Step 4: Repeat steps 2 and 3 until the model has been trained on the task sequence;

[0029] Step 5: Collect the accuracy rates calculated in Step 3 to form an accuracy matrix, and calculate the average accuracy rate of each row in the lower triangular part of the accuracy matrix to obtain the average accuracy vector shown in Equation (1). ;

[0030] Phase Two:

[0031] Step 6: Calculate the knowledge weight matrix according to equation (2). ;

[0032] Step 7: Calculate the average incremental accuracy of the context knowledge enhancement proposed in this invention according to equation (3). ;

[0033] The patentable aspect of this invention lies in the second stage. Any use of the second stage falls within the protection scope of this invention.

[0034] The following section compares different evaluation metrics using specific data. Let's assume the accuracy matrix obtained in step 5 is as shown in the first six columns of the table below:

[0035] ;

[0036] Then the average accuracy vector in equation (1) As shown in column 7 of the table above, the commonly used metric for average incremental accuracy is the average accuracy vector. The average of each accuracy rate is calculated as follows: . To calculate the average incremental accuracy of the contextual knowledge enhancement proposed in this invention, the knowledge weight matrix must first be calculated according to equation (2). (Pick )as follows:

[0037]

[0038]

[0039] ;

[0040] According to equation (3), the average incremental accuracy of contextual knowledge reinforcement can be calculated as follows:

[0041] .

Claims

1. An evaluation method for continuous learning of multi-source heterogeneous data based on contextual knowledge reinforcement, the method comprising: Step 1: Initialize a multi-source heterogeneous data processing network Its input data includes image and text modalities, and its structure includes an image data encoder. Text data encoder Modal interaction structure and classifier network Text data input text data encoder Image data input image data encoder Text data encoder and image data encoder Output-input modal interaction structure Modal interaction structure The multivariate heterogeneous data fusion feature is output, and the multivariate heterogeneous data fusion feature is input into the classifier, which outputs the category. Step 2: Process the dataset Preprocessing is performed to divide it into training sets. and test set The training set is divided into multiple subsets. The test set is divided according to the same criteria as the training set, resulting in... After this division, each subset of the training set and the subset of the test set constitute a dataset for a task. Multiple tasks sequentially fed into a multi-source heterogeneous data processing network constitute a task sequence. ; Indicates the total number of tasks; Step 3: In the task During training, data is used Training data in Training a multi-source heterogeneous data processing network, based on training data Data from each batch of samples The loss is calculated, and the parameters of the multi-source heterogeneous data processing network are optimized by backpropagating gradient updates. Step 4: In the task After training is completed, test data from all previously learned tasks will be used. The above tests the multi-source heterogeneous data processing network trained so far. And calculate the relevant accuracy, and obtain ; Step 5: Repeat steps 3 and 4 until the task sequence is complete. On-the-job training; Step 6: Collect the accuracy rates calculated in Step 4 to obtain the accuracy matrix; The average accuracy of the model across all tasks learned at that time is obtained by averaging the lower triangular portion of the accuracy matrix. This average accuracy is denoted as: ； Step 7: Based on the human forgetting curve ,in For the probability of recall, For the strength of memory traces, Set a knowledge weight matrix for time. ; Step 8: Calculate the average incremental accuracy of contextual knowledge reinforcement: 。 2. The evaluation method for continuous learning of multi-source heterogeneous data based on contextual knowledge reinforcement as described in claim 1, characterized in that, In step 2, the training set is divided into multiple subsets based on task differences, domain differences, or category differences.

3. The evaluation method for continuous learning of multi-source heterogeneous data based on contextual knowledge reinforcement as described in claim 1, characterized in that, The accuracy matrix in step 6 is as follows: .

4. The evaluation method for continuous learning of multi-source heterogeneous data based on contextual knowledge reinforcement as described in claim 1, characterized in that, In step 7: ; in, It is a natural constant. Indicates the task number.

Images