RESUMEN
Advances in deep learning have shown great promise towards the application of performing high-accuracy Electroencephalography (EEG) signal classification in a variety of tasks. However, many EEG-based datasets are often plagued by the issue of high inter-subject signal variability. Robust deep learning models are notoriously difficult to train under such scenarios, often leading to subpar or widely varying performance across subjects under the leave-one-subject-out paradigm. Recently, the model agnostic meta-learning framework was introduced as a way to increase the model's ability to generalize towards new tasks. While the original framework focused on task-based meta-learning, this research aims to show that the meta-learning methodology can be modified towards subject-based signal classification while maintaining the same task objectives and achieve state-of-the-art performance. Namely, we propose the novel implementation of a few/zero-shot subject-independent meta-learning framework towards multi-class inner speech and binary class motor imagery classification. Compared to current subject-adaptive methods which utilize large number of labels from the target, the proposed framework shows its effectiveness in training zero-calibration and few-shot models for subject-independent EEG classification. The proposed few/zero-shot subject-independent meta-learning mechanism performs well on both small and large datasets and achieves robust, generalized performance across subjects. The results obtained shows a significant improvement over the current state-of-the-art, with the binary class motor imagery achieving 88.70% and the accuracy of multi-class inner speech achieving an average of 31.15%. Codes will be made available to public upon publication.
Asunto(s)
Interfaces Cerebro-Computador , Humanos , Electroencefalografía/métodos , Calibración , Imaginación , AlgoritmosRESUMEN
In electroencephalography (EEG) classification paradigms, data from a target subject is often difficult to obtain, leading to difficulties in training a robust deep learning network. Transfer learning and their variations are effective tools in improving such models suffering from lack of data. However, many of the proposed variations and deep models often rely on a single assumed distribution to represent the latent features which may not scale well due to inter- and intra-subject variations in signals. This leads to significant instability in individual subject decoding performances. The presence of non-trivial domain differences between different sets of training or transfer learning data causes poorer model generalization towards the target subject. However, the detection of these domain differences is often difficult to perform due to the ill-defined nature of the EEG domain features. This study proposes a novel inference model, the Joint Embedding Variational Autoencoder, that offers conditionally tighter approximation of the estimated spatiotemporal feature distribution through the use of jointly optimised variational autoencoders to achieve optimizable data dependent inputs as an additional variable for improved overall model optimisation and scaling without sacrificing model tightness. To learn the variational bound, we show that maximising the marginal log-likelihood of only the second embedding section is required to achieve conditionally tighter lower bounds. Furthermore, we show that this model provides state-of-the-art EEG data reconstruction and deep feature extraction. The extracted domains of the EEG signals across each subject displays the rationale as to why there exists disparity between subjects' adaptation efficacy.
Asunto(s)
Aprendizaje Profundo , Electroencefalografía , HumanosRESUMEN
The commonly used fixed discrete Kalman filters (DKF) in neural decoders do not generalize well to the actual relationship between neuronal firing rates and movement intention. This is due to the underlying assumption that the neural activity is linearly related to the output state. They also face the issues of requiring large amount of training datasets to achieve a robust model and a degradation of decoding performance over time. In this paper, an adaptive adjustment is made to the conventional unscented Kalman filter (UKF) via intention estimation. This is done by incorporating a history of newly collected state parameters to develop a new set of model parameters. At each time point, a comparative weighted sum of old and new model parameters using matrix squared sums is used to update the neural decoding model parameters. The effectiveness of the resulting adaptive unscented Kalman filter (AUKF) is compared against the discrete Kalman filter and unscented Kalman filter-based algorithms. The results show that the proposed new algorithm provides higher decoding accuracy and stability while requiring less training data.