Generative Adversarial Network for iEEG compression

This thesis presents an algorithm to compressed multichannel signals which makes use of a Generative Adversarial Network for the first time. It is a modification of BPGAN by implementing LSTM neural networks instead of convolutional neural networks. BPGAN considers to select an optimal compressed representation that is also an element of a finite set of possible representations which allows a direct control of the value of compression ratio by selecting the cardinality of this set. The algorithm is tested on an intracranial electroencephalographic signal dataset, this has required some attentions because of the limited amount of information in literature about compressing iEEG and also because of the specific characteristics of this kind of signal. The BPGAN was implemented using a specific GAN architecture developed for synthetic EEG generation and the high spatial resolution of iEEG acquisition has introduced the necessity of applying the compression on separated group of channels that shows homogeneous characteristics. For this last reason a grouping algorithm that capture both linear and non-linear correlations is presented. Although some issues in the reconstruction quality associated with the discrete nature of the compression, this method reaches a state of art of 99.68% of relative compression ratio with 20.54% of mean PRD and 35.30 dB of mean PSNR.