TAFKAP: An improved method for probabilistic decoding of cortical activity

Cortical activity can be difficult to interpret. Neural responses to the same stimulus vary between presentations, due to random noise and other sources of variability. This unreliable relationship to external stimuli renders any pattern of activity open to a multitude of plausible interpretations. We have previously shown that this uncertainty in cortical stimulus representations can be characterized using a probabilistic decoding algorithm, which inverts a generative model of stimulus-evoked cortical responses. Here, we improve upon this method in two important ways, which both target the precision with which the generative model can be estimated from limited, noisy training data. We show that these improvements lead to considerably better estimation of the presented stimulus and its associated uncertainty. Estimates of the presented stimulus are recovered with an accuracy that exceeds that of standard decoding methods (SVMs), and in some cases even approaches the behavioral accuracy of human observers. Moreover, the uncertainty in the decoded probability distributions better characterizes the precision of cortical stimulus information from trial to trial.