Predictive coding has been one of the rare theories in neuroscience with bold, testable predictions at circuit level, and it’s been under scrutiny for years. It’s exciting to see recent experiments pushing it to its limits, hopefully leading to new directions. 🧠📈 - ThreadSky

shahabbakht.bsky.social • 158 days ago

Predictive coding has been one of the rare theories in neuroscience with bold, testable predictions at circuit level, and it’s been under scrutiny for years.

It’s exciting to see recent experiments pushing it to its limits, hopefully leading to new directions.

🧠📈

Reposted from Blake Richards

This paper may be very important:

www.biorxiv.org/content/10.1...

tl;dr: if you repeatedly give an animal a stimulus sequence XXXY, then throw in the occasional XXXX, there are large responses to the Y in XXXY, but not to the final X in XXXX, even though that's statistically "unexpected".

🧠📈 🧪

Comments

nafnlaus.bsky.social•155 days ago

Always nice to see more work on this, though the results seem to differ from eg. https://www.pnas.org/doi/abs/10.1073/pnas.1117807108 . Regardless, I'm not sure what implications exactly we're supposed to draw form L2/L3's prediction being overly simplistic. Is that not to be fully expected? Complexity emerges with depth.

shahabbakht.bsky.social•155 days ago

The difference between the two studies seems to be in the brain regions: https://bsky.app/profile/tyrellturing.bsky.social/post/3l5wewembyr2l

Although complexity emerges with depth, I expected the global predictions to influence lower areas via feedback, allowing lower representations to be contextualized by higher-level abstractions.

nafnlaus.bsky.social•155 days ago

Certainly one expects lower levels to be influenced by feedback, but there's only so much complexity that it's possible to store in such shallow layers. You expect more complexity than that? At L2/L3, visual cortex features for example are no more complex than orientation-invariant gratings/edges.

nafnlaus.bsky.social•155 days ago

Which is what we see in CNNs too. This for example is the conv2d1 level of Inception V1.

I don't think it's possible to get much more complexity than that in such shallow layers.

shahabbakht.bsky.social•155 days ago

I agree, the representational space can’t be too complex, but inference per noisy sample can benefit from context. E.g., inferring orientation in V1 for a noisy edge can benefit from the global shape of the object, which could appear as shape-aware modulation of orientation-tuned neurons.

tyrellturing.bsky.social•155 days ago

I would have thought that a traditional predictive coding model, trained on XXXY constantly, would then show a prediction error response to an XXXX in lower layer, yes. Why? Because this is still only a second or third-order prediction error (e.g. it's not a really abstract violation of statistics).

Comments

Posting Rules

Reply