Abstract
Brain signals carry cognitive information that can be relevant in downstream tasks, but what about eye-gaze? Although this can be estimated with eye-trackers, it can be very convenient in practice to do it without extra equipment. We consider the challenging tasks of fixation prediction and gaze estimation from electroencephalography (EEG) using deep learning models. We argue that there are three critical design criteria when designing neural architectures for EEG: (1) the spatial and temporal dimensions of the data, (2) the local vs global nature of the data processing, and (3) the overall structure and order with which the steps (1) and (2) are orchestrated. We propose two model architectures, based on Transformers and LSTMs, with different variants in this large design space, and compare them with recent state-of-the-art (SOTA) approaches under two constraints: reduced EEG signal length and reduced set of EEG channels. Our Transformer-based model outperforms the LSTM-only model, but it turns out to be more sensitive with short signal lengths and with less number of channels. Interestingly, our results are similar or slightly better than SOTA, and the models are trained from scratch (i.e., without pre-training or fine-tuning). Our findings provide useful insights for advancing in eye-from-EEG tasks.