Abstract
Representation of stimuli by neural ensembles and the correlation of neural activity with the behavioral choice are, in principle, two different computational problems, however, it is only the intersection between the two that is relevant for animal’s behavior. Here, we propose a decoding model that learns its decoding weights in the presence of the information on both the stimulus class and the future behavioral choice. We then test the model on trials that only differ in the choice and show that the choice can be read-out from the activity of populations in V1. The read-out model is a linear weighted sum of spikes, decoding the behavioral choice in single trials and without temporal averaging. The generalization of learning suggests that the representation of the stimulus class and of the behavioral choice have a non-zero intersection. In addition, we show how the spike timing is required for discrimination, that bursty neurons carry more information than non-bursty neurons and that neurons in the superficial layer are the most important for discrimination.
Highlights
Learning from both stimuli and behavioral choice generalizes to the representation of the choice alone.
With generalization of learning, the choice signal can be read-out from parallel spike trains in V1.
Correct attribution to binary coding pools and correct spike timing are necessary for the read-out.
Bursty neurons convey more information than non-bursty neurons. Discrimination is the strongest in the superficial layer of the cortex.