This paper shows that the earliest stages of visual processing in macaque monkeys are more dynamic than often assumed. Instead of a single, stage-like feedforward pass that hands a stable object code from one brain area to the next, the authors find multiple, different bouts of information exchange between mid-level and high-level visual areas within the first moments after a picture appears.

The researchers recorded local field potentials (LFPs) — a measure of nearby electrical activity — from many implanted electrode arrays across the ventral visual stream of two macaques. They used 16 (15 in the second animal) 64-channel Utah arrays placed in primary visual cortex (V1), visual area V4, and inferotemporal cortex (IT). The animals passively viewed 1,000 natural images drawn from 100 object categories while the scientists measured activity. LFPs were chosen because they had higher signal-to-noise in this data set.

To study how information moves between areas, the team built time-resolved linear models that predict activity in IT from earlier activity in V4 while accounting for conduction delays. These models revealed two separate V4→IT communication windows where V4 activity added predictive power. One was an “early” interaction (V4 at about 46 ms predicting IT at about 74 ms after stimulus onset) and a “late” interaction (V4 around 80 ms predicting IT around 120 ms). The two interactions carried different kinds of information: representational dissimilarity analyses (which compare how the pattern across stimuli changes) showed that the early and late predictions produced distinct representational geometries. In other words, IT seemed to read different messages from V4 at different times.

The authors compared these time-varying patterns to layers of an artificial neural network (ANN, a computer model trained to classify images). Early V4→IT predictions resembled low-level ANN representations, while the later predictions matched more abstract, category-like ANN layers. They also used recurrent neural network (RNN) decoders — models that take temporal sequences into account — and found that the temporal dynamics of the neural population carried categorical information beyond what was present in any single instant. In short, the evolving activity patterns themselves help encode object category.