It’s a short read, highlighting open questions about where and how feature-specific prediction errors are computed and relayed across the visual hierarchy.
Take a look!

direct.mit.edu/imag/article...

In our article, we discuss whether and how four accounts might explain these results:
(1) hierarchical predictive coding,
(2) feedback propagation of error signals,
(3) V1 as a comparator circuit for higher-level features,
(4) dendritic HPC.

Rather than focusing only on the magnitude of surprise, studies have begun to probe the content of prediction errors, showing that even early visual responses may primarily scale with high-level, rather than low-level, visual surprise.

Congratulations Peter! Amazing news and well deserved!

Thanks Juan!

If you’re interested in more details, check out the full paper:
doi.org/10.1016/j.is...

Taken together, our findings show that high-level visual predictions are rapidly integrated during perceptual inference, suggesting that the brain's predictive machinery is finely tuned to utilize expectations abstracted away from low-level sensory details to facilitate perception.

We also found a small decrease in neural responses by semantic (word-based) surprise. Notably, low-level visual surprise had no detectable effect, even though stimuli were predictable all the way down to the pixel level.

Then we turned to the key questions: When and what kind of surprise drives visually evoked responses?
Neural responses ~190 ms post-stimulus onset over parieto-occipital electrodes were selectively increased by high-level visual surprise!

As a sanity check, we first used RSA to show that the CNN and other models of interest (semantic and task models) robustly explained the EEG responses independent of surprise.

We investigated these questions using EEG and a visual CNN. Participants viewed object images that were probabilistically predicted by preceding cues. We then quantified surprise trial-by-trial at low-levels (early CNN layers) and high-levels (late CNN layers) of visual feature abstraction.

Predictive processing theories propose that the brain continuously generates predictions about incoming sensory input.
But what exactly does the brain predict? Low-level (edges, contrasts) and/or high-level visual features (textures, objects)?
And when do these predictions shape neural responses?

If you are interested in more details check out the preprint here:
www.biorxiv.org/content/10.1...

Taken together, our findings demonstrate that high-level visual predictions are rapidly integrated during perceptual inference. This suggests that the brain's predictive machinery is finely tuned to utilize expectations abstracted away from low-level sensory details, likely to facilitate perception.