Conference on Cognitive Computational Neuroscience

The early visual system is believed to build visual representations based on a dynamic code that is flexibly adapted to suit the behavioral goals of the observer. Efficient coding theories posit that redundant input information is compressed. However, the evidence in support of that process is largely descriptive. Here, we propose that information reduction can be achieved through task-relevant dynamic sparse-distributed coding. Using high-density EEG signals from humans engaged in different scene understanding tasks, we developed a brain-supervised sparse coding network to learn visual filters based on EEG variance over time. Stimuli were reconstructed based on filters that were sparsified according to the brain variance across time. The results show that information reduction via sparsification does not follow a simple ‘many-to-few’ operation, but instead differentially compresses and expands information over time in a task-dependent manner. This suggests that efficient coding may not result in a terminal state of compression, but instead operates to yield a goal-oriented state of dynamic equilibrium.