Unsupervised Learning for Exploring Hidden Structures in Self-Talk

Innerspeech decoding from EEG data holds significant importance due to its potential to revolutionize human-machine interaction and communication systems. Leveraging the power of temporal shift-invariant sparse coding, this study explores the unsupervised learning of inner-speech patterns using EEG, a prominent modality in body sensor networks. By analyzing EEG data, we investigate the characteristics of sparse code activities to distinguish between different classes and conditions. The results showcase the effectiveness of the unsupervised model, emphasizing its potential for accurate inner speech decoding without the need for explicit class labels. Furthermore, we assess the significance of sparse code activities using an ANOVA test, providing statistical evidence of their discriminative power across different classes and conditions. To assess the discriminatory power of dictionaries, we compare the results of Multilayer Perceptron (MLP) and Convolutional Neural Networks (CNN) classifiers on both raw data and dictionary outputs. The findings demonstrate that the accuracy does not decrease when employing the dictionary approach, showcasing its effectiveness in inner speech decoding. This research significantly contributes to the field of signal processing and body sensor networks, paving the way for advancements in innerspeech decoding and its potential applications across a diverse range of domains.