Examining the Size of the Latent Space of Convolutional Variational Autoencoders Trained With Spectral Topographic Maps of EEG Frequency Bands

Dimensionality reduction and the automatic learning of key features from electroencephalographic (EEG) signals have always been challenging tasks. Variational autoencoders (VAEs) have been used for EEG data generation and augmentation, denoising, and automatic feature extraction. However, investigations of the optimal shape of their latent space have been neglected. This research tried to understand the minimal size of the latent space of convolutional VAEs, trained with spectral topographic EEG head maps of different frequency bands, that leads to the maximum reconstruction capacity of the input and maximum utility for classification tasks. Head maps are generated employing a sliding window technique with a 125ms shift. Person-specific convolutional VAEs are trained to learn latent spaces of varying dimensions while a dense neural network is trained to investigate their utility on a classification task. The empirical results suggest that when VAEs are deployed on spectral topographic maps with shape $32\times 32$ , deployed for 32 electrodes from 2 seconds cerebral activity, they were capable of reducing the input up to almost 99%, with a latent space of 28 means and standard deviations. This did not compromise the salient information, as confirmed by a structural similarity index, and mean squared error between the input and reconstructed maps. Additionally, along the 28 means maximized the utility of latent spaces in the classification task, with an average 0.93% accuracy. This study contributes to the body of knowledge by offering a pipeline for effective dimensionality reduction of EEG data by employing convolutional variational autoencoders.