Conference on Cognitive Computational Neuroscience

Extensive literature has characterized the behavioral and neural correlates of attention with spatial probabilistic cueing, also known as ``Posner'' cueing. Yet, it has been argued that Posner cueing tasks conflate the effects of attention and expectation: in such tasks event expectation carries information about where attention must be directed. We address this question with transfer learning applied to a deep CNN(convolutional neural network) based decoder. First, we developed a novel, embedding-filtered CNN architecture (EFECTS) to decode task conditions from single-trial electroencephalography data(19 participants, 16,843 trials). We then trained two distinct models to decode orthogonal spatial attention and expectation cues in a dual cueing task. EFECTS achieved accuracies of >85% and >65% at decoding attention and expectation cues, respectively, exceeding competing models(EEGnet, HTnet) by a significant margin: >10%. Saliency maps revealed distinct spectro-temporal neural signatures of spatial attention and expectation. Finally, we transferred each model(attention, expectation) to a Posner cueing task, with similar task sequence. Both models successfully inferred the Posner cued location on a single-trial level. The results show that Posner cueing induces neural signatures of both spatial attention and expectation. Our embedding-filtered deep CNN provides a parsimonious framework, surpassing the state-of-the-art, for decoding cognitive states from noisy EEG recordings.