Decoding of EEG Signals Using Deep Long Short-Term Memory Network in Face Recognition Task

The paper proposes a novel approach to classify the human memory response involved in the face recognition task by the utilization of event related potentials. Electroencephalographic signals are acquired when a subject engages himself/herself in familiar or unfamiliar face recognition tasks. The signals are analyzed through source Iocalization using eLORETA and artifact removal by ICA from a set of channels corresponding to those selected sources, with an ultimate aim to classify the EEG responses of familiar and unfamiliar faces. The EEG responses of the two different classes (familiar and unfamiliar face recognition)are distinguished by analyzing the Event Related Potential signals that reveal the existence of large N250 and P600 signals during familiar face recognition.The paper introduces a novel LSTM classifier network which is designed to classify the ERP signals to fulfill the prime objective of this work. The first layer of the novel LSTM network evaluates the spatial and local temporal correlations between the obtained samples of local EEG time-windows. The second layer of this network models the temporal correlations between the time-windows. An attention mechanism has been introduced in each layer of the proposed model to compute the contribution of each EEG time-window in face recognition task. Performance analysis reveals that the proposed LSTM classifier with attention mechanism outperforms the efficiency of the conventional LSTM and other classifiers with a significantly large margin. Moreover, source Iocalization using eLORETA shows the involvement of inferior temporal and frontal lobes during familiar face recognition and pre-frontal lobe during unfamiliar face recognition. Thus, the present research outcome can be used in criminal investigation, where meticulous differentiation of familiar and unfamiliar face detection by criminals can be performed from their acquired brain responses.