Using intersection information to map stimulus information transfer within neural networks

Analytical tools that estimate the directed information flow between simultaneously recorded neural populations, such as directed information or Granger causality, typically focus on measuring how much information is exchanged between such populations. However, understanding how sensory information is processed through the brain and how it is used to generate behaviors requires estimating specifically the amount of stimulus information that is transmitted. Here we use the concept of intersection information to make progress on how to perform this measure. We develop the concept of transmitted intersection information, which measures how much of the stimulus information present in one population at a certain time is transmitted to a second population at a later time. We show that this measure of stimulus-specific information transfer has several appealing properties, such as being non-negative, and being bounded by the amount of stimulus information present in each of the two populations and by the total amount of information transmitted between the two populations. Applying this measure to simulated neurons or pools of neurons connected by feed-forward synapses, we show that it can discern cases when the information transmitted from one population to another is about specific stimulus features encoded by the sending population from cases in which the information transmitted is not about the stimuli. We also show that this measure has a good statistical sensitivity from trial numbers that can be collected in real data. Our results highlight the promise of using the concept of intersection information to map stimulus-specific information transfer across neural populations.