EEG decoding method based on unsupervised individual continuous learning

Abstract

The invention discloses an EEG decoding method based on unsupervised individual continuous learning, which improves the performance of a model on unknown subjects on the basis of improving the performance of incremental individuals by designing an unsupervised individual continuous learning framework and training an incremental model, and further solves the problem of generalization of the EEG model on the unknown subjects. According to the method, a self-supervised contrast learning method is utilized to generate a high-quality incremental individual pseudo tag, a cache region based on dynamic confidence is utilized to select and play back a mixed true and false tag for fine tuning of a sample, and then a cross-round alignment algorithm is utilized to reduce potential drift of model distribution along with continuous learning and a feature space. Therefore, the generalization of the incremental model for unknown subjects is improved while the performance of the incremental individual is improved.

Description

Technical Field

[0001] The present invention belongs to the technical field of EEG decoding, and in particular relates to an EEG decoding method based on unsupervised individual continuous learning. Background Art

[0002] Electroencephalography (EEG) is a non-invasive brain-computer interface (BCI) technology that records the brain's electrical activity through electrodes placed on the scalp. Due to its non-invasiveness and relatively high temporal resolution, EEG plays an important role in daily life and is widely used in practice, especially in clinical fields such as sleep staging, emotion recognition, motor imagery, and disease diagnosis. However, existing EEG-related models perform poorly in practical applications; in actual scenarios, new unknown individuals appear every day, and there are significant individual differences between different individuals (such as physiological structure and physical characteristics). Therefore, these EEG models trained based on fixed data sets are difficult to generalize to new and unseen individuals, resulting in poor performance on unknown individuals. Therefore, the present invention application attempts to solve the limitations of existing EEG models in reality in order to promote the development of EEG in practical applications.

[0003] This problem can be attributed to continuous learning (CL), in which the model can learn and improve by continuously absorbing new knowledge. The main challenge of continuous learning is to balance stability and plasticity, where plasticity refers to the ability of the model to adapt to newly emerged individuals, while stability refers to the model's ability to generalize to seen and unseen individuals. Inspired by continuous learning and considering the lack of true labels for newly emerged individuals, we propose a novel unsupervised individual continuous learning (UICL) paradigm to address the problem of EEG tasks in practical applications. Considering individual differences, we regard each individual as an independent individual domain in our study; in this paradigm, the incremental model needs to continuously adapt to multiple individual target domains in turn, while constantly absorbing new knowledge from incremental individuals and self-improving, so that the model can eventually generalize to all unknown individuals.

[0004] We have two main goals: (1) better plasticity, the model can adapt well to each new individual in the continuous individual stream; (2) better stability, the model can continuously learn new knowledge in the continuous learning process, so that it can generalize better to those unknown individuals. However, it is not easy to make the model achieve better plasticity and better stability in the continuous learning process, mainly for the following three reasons: First, due to the individual differences in the continuous individual stream, there is a continuous domain shift between the source domain and the individual target domain distribution; second, in the continuous learning process, the model may overfit some individual target domains in pursuit of better plasticity, especially when adapting to some abnormal individuals, the model performance may drop sharply, and even cannot be recovered in the subsequent continuous adaptation process; third, it is a challenging task to make the model improve its generalization ability to all unseen individuals while adapting to new individuals.

[0005] Although there are studies that attempt to solve similar problems, they are less adaptable in practical applications. Their research mainly targets target domains with small changes and solves the stability-plasticity problem in scenarios such as object detection and image classification, but usually assumes that the domain changes between consecutive batches are small and studies are conducted at the sample level. However, the actual scenario is completely different. There is a continuous flow of a large number of new individuals in the actual scenario, and the changes between individual domains are significant; therefore, in practical applications, research based on continuous EEG decoding should be conducted at the individual level (continuous learning based on individuals) rather than the sample level. Summary of the invention

[0006] In view of the above, the present invention provides an EEG decoding method based on unsupervised individual continuous learning, by designing an unsupervised individual continuous learning framework and training a deep neural network incremental model to continuously learn new knowledge from incremental individuals, thereby improving the performance of the model on incremental individuals and continuously improving the generalization of the model, thereby solving the problem of insufficient generalization of the EEG model in the face of unknown individuals.

[0007] An EEG decoding method based on unsupervised individual continuous learning includes the following steps: (1) Obtain datasets based on different EEG downstream tasks, where each individual is considered as an independent individual domain; (2) Preprocess the EEG signals in the dataset for model training; (3) A small number of labeled individual domains are selected from the dataset as source domains, and the remaining individual domains are divided into target domains and generalized domains; (4) Construct an incremental learning model and use the source domain data to pre-train the incremental learning model for subsequent unsupervised individual continuous EEG decoding; use the pre-trained incremental learning model to perform continuous learning on individuals in the target domain and test the model on the generalization domain.

[0008] Furthermore, the data sets obtained in step (1) are from three different research institutions, namely, ISRUC, FACED, and Physionet-MI, three public data sets. Each data set contains more than 100 subjects. Due to significant differences between individuals, the EEG signal distribution of each individual is independent, so each individual is regarded as an independent individual domain.

[0009] Furthermore, the number of channels and specific positions of the EEG signals contained in different data sets in step (2) are different. For the ISRUC data set, the 8 channels contained therein are F3-A2, C3-A2, O1-A2, F4-A1, C4-A1, O2-A1, E1-M2, and E2-M1; for the FACED data set, all 32 channels contained therein are used; and for the Physionet-MI data set, all 64 channels contained therein are used; and then the EEG signals of the selected channels are pre-processed by resampling at 100 Hz.

[0010] Furthermore, the incremental learning model in step (3) includes a convolutional neural network, a Transformer and a classifier, wherein the convolutional neural network is used to extract local features within each segment of the input EEG signal, and the Transformer extracts context features between the input EEG signal segments based on the local features. The classifier is composed of a fully connected layer and a Softmax function, and based on different EEG downstream tasks, outputs the predicted probability that each segment belongs to each category in different tasks.

[0011] Furthermore, in step (4), the pre-trained incremental learning model is used to continuously learn the individuals in the target domain. The specific process is as follows: 4.1 Establish a cache area to store source domain data and incremental individual data; 4.2 Copy the existing model as the guide model (the initial existing model is a pre-trained incremental learning model), select an incremental individual from the target domain, and use the incremental individual data to fine-tune the guide model through a self-supervised contrastive learning method; 4.3 Input the incremental individual data into the fine-tuned bootstrap model to obtain the pseudo labels of the incremental individual data; 4.4 Using the replay data, incremental individual data and their pseudo-labels in the cache, an alignment algorithm based on cross-training rounds is used to adaptively learn the existing model, so that the new model can adapt to the data distribution of incremental individuals, thereby completing an incremental individual domain adaptation; 4.5 Input the incremental individual data into the new model to obtain new pseudo labels for the incremental individual data, and select the incremental individual data with high confidence (by threshold comparison) and its new pseudo labels and store them in the cache; 4.6 When performing the next incremental individual domain adaptation, use the new model obtained in step 4.4 as the existing model and return to execute step 4.2.

[0012] Furthermore, the cache is divided into a true label cache and a pseudo label cache, wherein the true label cache is used to store source domain data with real labels, and the pseudo label cache is used to store incremental individual data and its pseudo labels; the playback data selectively mixes data from different buffers according to a certain sampling ratio for adaptive learning of the existing model, wherein the data selected from the true label cache accounts for 80% of the playback data, and the data selected from the pseudo label cache accounts for 20% of the playback data.

[0013] Furthermore, in step 4.4, an alignment algorithm based on cross-training rounds is used to adaptively learn the existing model, that is, different total loss functions L are used in different training rounds. overall , in odd rounds L overall =L C ; In even rounds L overall =L C +L KL , where L C is the loss function fine-tuned by incremental individual pseudo-labels and replay data, L KL is the loss function for cross-round alignment.

[0014] Furthermore, the loss function L C The expression is as follows:

[0015] in: represents the probability that the playback data sample is predicted by the model as the cth category, y Bc Indicates the probability that the corresponding label of the playback data sample is the cth category, represents the probability that the incremental individual data sample is predicted by the model as the cth category, It represents the probability that the pseudo label corresponding to the incremental individual data sample is the cth category, α is the weighting coefficient, and c is the index number of the category.

[0016] Furthermore, the loss function L KL The expression is as follows:

[0017] in: express and The KL divergence between represents the feature distribution of the playback data on the model in the e-th training round, Represents the feature distribution of the playback data on the model in the e+2th training round, where e is a positive even number including 0.

[0018] Based on the above technical scheme, the present invention uses a small amount of annotated source domain data to pre-train the incremental model, uses self-supervised contrastive learning to fit the domain distribution of incremental individuals, generates high-quality pseudo labels for fine-tuning, and then uses a hybrid replay strategy to output replay samples that take into account both sample authenticity and diversity. Finally, a cross-round alignment algorithm is used to align the feature distribution space of replay samples in different training stages to avoid model drift, improve the performance of the model on incremental individuals, and take into account the generalization of the incremental model on unknown individuals. BRIEF DESCRIPTION OF THE DRAWINGS

[0019] Figure 1 It is a flow chart of the EEG decoding method based on unsupervised individual continuous learning of the present invention.

[0020] Figure 2 This is an overall schematic diagram of the EEG decoding framework based on unsupervised individual continuous learning of the present invention.

[0021] Figure 3 It is a flow chart of the continuous learning stage of the present invention. DETAILED DESCRIPTION

[0022] In order to describe the present invention more specifically, the technical solution of the present invention is described in detail below in conjunction with the accompanying drawings and specific implementation methods.

[0023] like Figure 1 As shown, the EEG decoding method of the present invention is based on unsupervised individual continuous learning, and the specific steps are as follows: (1) We selected three datasets of different EEG downstream tasks from different research institutions, namely ISRUC, FACED, and Physionet-MI, each of which contains more than 100 subjects. Since there are significant differences between individuals, the EEG signal distribution of each individual is independent, so each individual can be regarded as an independent individual domain.

[0024] ISRUC is a sleep dataset from the Institute for Systems and Robotics at Columbia University, which contains polysomnographic records, including EEG, EOG, EMG and other signals, as well as expert annotations of sleep stages; FACED is a fine-grained affective computing EEG dataset from Tsinghua University, which contains 32-channel EEG signals of 123 subjects, recorded while watching 28 emotion-inducing videos; Physionet-MI is an EEG dataset provided by PhysioNet, which contains multi-channel EEG signals, recording EEG signal data when subjects perform or imagine specific motor tasks (such as left hand, right hand, and foot movements).

[0025] (2) The number and specific locations of the channels of physiological electrical signals contained in different datasets are different. For the ISRUC dataset, six EEG signal channels F3-A2, C3-A2, O1-A2, F4-A1, C4-A1, O2-A1 and two EOG signal channels E1-M2 and E2-M1 are used, and the selected channels are pre-processed by 100 Hz resampling; for the FACED dataset, all 32 channels are used; and for the Physionet-MI dataset, all 64 channels are used.

[0026] (3) Based on the idea of ​​continuous learning, the present invention adopts Figure 2 The EEG decoding framework based on unsupervised individual continuous learning is shown in the figure for continuous individual EEG decoding. The framework generates high-quality individual pseudo labels with the help of self-supervised contrastive learning methods for fine-tuning the incremental model. It reduces the drift of the model distribution during continuous learning through a dynamic confidence-based cache and an algorithm based on the idea of ​​cross-round alignment, thereby improving the performance of the model on incremental individuals while ensuring the stability of the model.

[0027] The incremental model consists of three parts: a convolutional neural network encoder, a Transformer encoder, and a classifier. The convolutional neural network is used to extract local features within each EEG segment in the input signal. The Transformer extracts contextual features between input EEG sequence signals based on the EEG segment features. The classifier consists of a fully connected layer and a softmax function. Based on different EEG downstream tasks, it outputs the predicted probability that each EEG segment belongs to different categories in different tasks.

[0028] The self-supervised contrastive learning method adopts a method based on temporal coding prediction. Specifically, we first start from the latest incremental model M i-1 A guidance model is copied from the source, and contrastive coding learning is used on the guidance model. According to the temporal nature of the EEG sequence, the early EEG sequence is used to predict the later sequence. After fine-tuning the guidance model, an incremental individual pseudo label with high confidence is output. , and then use the generated pseudo labels to fine-tune the incremental model. In this step, the guidance model is only responsible for generating pseudo labels. , and does not participate in the fine-tuning of the subsequent incremental model.

[0029] The dynamic confidence-based buffer consists of the true label buffer S true and pseudo-label buffer S pseudo Specifically, the samples cached in the true label buffer consist of sample-label pairs that are truly annotated in the source domain. After each individual incremental learning, the pseudo label buffer is filled with sample-pseudo label pairs with high confidence generated by the incremental model on the incremental individuals. In general, the number of samples in the true label buffer remains unchanged, while the number of samples in the pseudo label buffer increases with the continuous learning process. Each playback selectively mixes and replays data from different buffers according to the sampling ratio for incremental model fine-tuning. The true label buffer accounts for 80% of the playback samples, and the pseudo label buffer accounts for 20%, which not only ensures the authenticity of the playback samples, but also provides diversity of the playback samples.

[0030] Based on the alignment algorithm across training rounds, by aligning the KL divergence of the playback data of different training rounds in the feature space, the drift of the model distribution in the feature space is reduced, which can effectively prevent the model from being overly affected by new individuals, especially abnormal individuals, during the continuous learning process.

[0031] Incremental model from M i-1 Iterate to M i The specific steps are as follows: First, from the latest incremental model M i-1 Copy a guided model from the source code, use contrastive coding learning on the guided model, fine-tune the guided model and output incremental individual pseudo labels with high confidence Next, replay the sample data and labels from the buffer, denoted as X B and B , and the sample-pseudo label pair of the newly added individual and Mixed input incremental model M i-1 A total of 10 rounds of training are performed. In the odd-numbered training rounds, the feature distribution of the model output in the current training round is calculated based on the newly added individual sample data. And the classification results , only using cross entropy loss and Update model parameters; in even-numbered training rounds, use the cross-training round alignment algorithm to align the playback data in the current training round model Output feature distribution The model is trained with the replay data in the first two epochs Output feature distribution F B And the classification results , using cross entropy loss and And KL divergence loss L KL Update the model parameters together. After completing 10 rounds of training, the incremental model is i-1 Iterate to M i , and then the incremental individual samples Input to M i In the above example, we get the model M i For the predicted pseudo labels of incremental individual samples, retain the high-quality sample-pseudo label pairs and update them to the cache area.

[0032] (4) First, a small part of the labeled individual domain is selected as the source domain, and the remaining individual domain is divided into incremental individual domain and generalization domain; on the source domain data, the incremental learning model is pre-trained using the training set and validation set as the model selection strategy. Figure 3 The continuous learning phase process shown in the figure performs continuous individual EEG decoding. Specifically, the pre-trained incremental model M0 performs unsupervised continuous learning on each individual domain in the target domain in turn, continuously adapts to the data distribution of each incremental individual domain, learns the EEG signal characteristics of different incremental individuals, and iterates to M1, M2, M3... Each time the model completes incremental individual domain adaptation, the performance of the model on the incremental individual is evaluated, and the generalization of the incremental model is tested on the generalization domain, and the performance of the incremental model on unknown subjects on different EEG tasks is evaluated; it should be emphasized that the incremental individual domain is completely unknown and unsupervised in the continuous learning process.

[0033] After the data set is divided, the source domain data is input into the incremental model for pre-training. On the three selected data sets, the pre-training process only includes classification tasks, and the loss function of the pre-training classification task is L pretrain as follows:

[0034] Where: c represents the number of categories of tasks in the EEG downstream task, represents the probability that the EEG sample is predicted to be the cth category during training, y c express The corresponding true probability.

[0035] After pre-training, the incremental model needs to go through the following steps to complete continuous learning for each new individual: first, copy a guide model from the latest incremental model, use contrastive coding learning on the guide model, and use the early EEG sequence to predict the later sequence according to the temporal nature of the EEG sequence. After fine-tuning the guide model, output the pseudo-label of the incremental individual with high confidence. The loss function L of self-supervised contrastive learning is used on the guide model. ssl As shown below:

[0036] Where: t represents the signal after the tth time step in the EEG sequence that needs to be predicted, h t represents the intermediate feature corresponding to the tth time step in the EEG sequence, f k represents the kth linear layer, c t Represents the context features corresponding to the previous t time steps.

[0037] After guiding the model to generate pseudo labels, the incremental model needs to be fine-tuned together with the data replayed from the buffer. The replayed data is output by the dynamic confidence buffer using a mixed replay strategy. Each replay is selectively mixed with the real label buffer S according to the sampling ratio. true and pseudo-label buffer S pseudo The data is used for incremental model fine-tuning as described below:

[0038] The model is fine-tuned using incremental individual pseudo-labels and mixed playback data, and its loss function L C The expression is as follows:

[0039] in: represents the mixed playback data fine-tuning loss function, represents the incremental individual pseudo-label fine-tuning loss function, represents the probability that the replayed sample is predicted to be the cth category, y Bc express The corresponding true probability is represents the probability that the incremental individual sample is predicted to be the cth category, express The true probability of the corresponding pseudo label, α is the weighting coefficient.

[0040] For individual incremental learning training, after every two training rounds, a cross-round alignment of the playback sample feature distribution is performed, and its loss function L KL The expression is as follows:

[0041] Where: M i-1 represents the current i-1th incremental model, θ represents the incremental model parameters, X B Represents samples played back from the buffer. represents the distribution of the incremental model at the e-th training round, represents the distribution of the incremental model at the e+2th training round, D KLIndicates calculating the KL divergence between different distributions.

[0042] Therefore, the total loss function L in the continuous learning training process overall As shown below:

[0043] Where: L C is the loss function for incremental individual pseudo-label fine-tuning and mixed replay data fine-tuning, L KL is the loss function for cross-round alignment, and e represents the e-th training round in the individual continuous learning training process.

[0044] Each time the model completes domain adaptation for an incremental individual, the data of the current incremental individual is re-input, the model outputs the corresponding pseudo-label, and updates the sample-pseudo-label pairs with predicted probabilities greater than the threshold into the pseudo-label cache.

[0045] (5) If Figure 3 As shown, each incremental individual is learned, the performance of the model on the incremental individual is verified, and the performance of the model on the generalization set is verified, and the accuracy (ACC) and Macro-F1 score (MF1) of the two are calculated. When the model is continuously learned, the average performance of all incremental individuals is counted, and the stability change indicators AAA and AAF1 of the brightening model on the generalization set during the continuous learning process are calculated. The calculation method of the stability evaluation indicators AAA and AAF1 is as follows:

[0046]

[0047] Where: i represents the i-th increment individual, N G represents the number of individuals in the generalization domain, and They respectively represent the predicted label and the true label of the incremental model in the generalization domain after adapting to the i-th incremental individual.

[0048] A comparison with other existing methods is shown in Table 1. It can be seen from the table that compared with other existing methods, the method of the present invention not only significantly improves the performance of the incremental model on the incremental individuals, but also significantly improves the generalization performance of the incremental model on the generalization set.

[0049] Table 1

[0050] The above description of the embodiments is to facilitate the understanding and application of the present invention by those skilled in the art. It is obvious that those skilled in the art can easily make various modifications to the above embodiments and apply the general principles described herein to other embodiments without creative work. Therefore, the present invention is not limited to the above embodiments, and improvements and modifications made by those skilled in the art based on the disclosure of the present invention should be within the scope of protection of the present invention.

Claims

1. An EEG decoding method based on unsupervised individual continuous learning, comprising the following steps: (1) Obtain datasets based on different EEG downstream tasks, where each individual is considered as an independent individual domain; (2) Preprocess the EEG signals in the dataset for model training; (3) A small number of labeled individual domains are selected from the dataset as source domains, and the remaining individual domains are divided into target domains and generalized domains; (4) Construct an incremental learning model and use the source domain data to pre-train the incremental learning model for subsequent unsupervised individual continuous EEG decoding; use the pre-trained incremental learning model to perform continuous learning on individuals in the target domain and test the model on the generalization domain.

2. The EEG decoding method based on unsupervised individual continuous learning according to claim 1, characterized in that: The data sets obtained in step (1) are from three different research institutions, namely, ISRUC, FACED, and Physionet-MI, three public data sets. Each data set contains more than 100 subjects. Due to significant differences between individuals, the EEG signal distribution of each individual is independent, so each individual is regarded as an independent individual domain.

3. The EEG decoding method based on unsupervised individual continuous learning according to claim 2, characterized in that: The number of channels and specific positions of the EEG signals contained in different data sets in step (2) are different. For the ISRUC data set, the 8 channels contained therein are F3-A2, C3-A2, O1-A2, F4-A1, C4-A1, O2-A1, E1-M2, and E2-M1; for the FACED data set, all 32 channels contained therein are used; for the Physionet-MI data set, all 64 channels contained therein are used; and then the EEG signals of the selected channels are pre-processed by resampling at 100 Hz.

4. The EEG decoding method based on unsupervised individual continuous learning according to claim 1, characterized in that: The incremental learning model in step (3) includes a convolutional neural network, a Transformer and a classifier, wherein the convolutional neural network is used to extract local features within each segment of the input EEG signal, and the Transformer extracts context features between the input EEG signal segments based on the local features. The classifier is composed of a fully connected layer and a Softmax function, and based on different EEG downstream tasks, outputs the predicted probability that each segment belongs to each category in different tasks.

5. The EEG decoding method based on unsupervised individual continuous learning according to claim 1, characterized in that: In step (4), the pre-trained incremental learning model is used to continuously learn the individuals in the target domain. The specific process is as follows: 4.1 Establish a cache area to store source domain data and incremental individual data; 4.2 Copy the existing model as the guide model, randomly select an incremental individual from the target domain, and use the incremental individual data to fine-tune the guide model through a self-supervised contrastive learning method; 4.3 Input the incremental individual data into the fine-tuned bootstrap model to obtain the pseudo labels of the incremental individual data; 4.4 Using the replay data, incremental individual data and their pseudo-labels in the cache, an alignment algorithm based on cross-training rounds is used to adaptively learn the existing model, so that the new model can adapt to the data distribution of incremental individuals, thereby completing an incremental individual domain adaptation; 4.5 Input the incremental individual data into the new model to obtain new pseudo labels for the incremental individual data, and select the incremental individual data with high confidence and its new pseudo labels and store them in the cache; 4.6 When performing the next incremental individual domain adaptation, use the new model obtained in step 4.4 as the existing model and return to execute step 4.

2.

6. The EEG decoding method based on unsupervised individual continuous learning according to claim 5, characterized in that: The cache area is divided into a true label cache area and a pseudo label cache area, wherein the true label cache area is used to store source domain data with real labels, and the pseudo label cache area is used to store incremental individual data and its pseudo labels; the playback data selectively mixes data from different buffers according to a certain sampling ratio for adaptive learning of the existing model, wherein the data selected from the true label cache area accounts for 80% of the playback data, and the data selected from the pseudo label cache area accounts for 20% of the playback data.

7. The EEG decoding method based on unsupervised individual continuous learning according to claim 5, characterized in that: In step 4.4, an alignment algorithm based on cross-training rounds is used to adaptively learn the existing model, that is, different total loss functions L are used in different training rounds. overall , in odd rounds L overall =L C ; In even rounds L overall =L C +L KL , where L C is the loss function fine-tuned by incremental individual pseudo-labels and replay data, L KL is the loss function for cross-round alignment.

8. The EEG decoding method based on unsupervised individual continuous learning according to claim 7, characterized in that: The loss function L C The expression is as follows: ; in: represents the probability that the playback data sample is predicted by the model as the cth category, y Bc Indicates the probability that the corresponding label of the playback data sample is the cth category, represents the probability that the incremental individual data sample is predicted by the model as the cth category, It represents the probability that the pseudo label corresponding to the incremental individual data sample is the cth category, α is the weighting coefficient, and c is the index number of the category.

9. The EEG decoding method based on unsupervised individual continuous learning according to claim 7, characterized in that: The loss function L KL The expression is as follows: ; in: express and The KL divergence between represents the feature distribution of the playback data on the model in the e-th training round, Represents the feature distribution of the playback data on the model in the e+2th training round, where e is a positive even number including 0.

Images