Always include some stimuli that are permissively licensed so they can be used as examples! E.g. we have a video stimulus set that's mostly Pixar short films, but also includes a segment from the Blender movie Sintel (en.wikipedia.org/wiki/Sintel), which is licensed CC-BY.

Someone on the aggies subreddit posted the course description and it seems pretty thorough! www.reddit.com/r/aggies/s/C...

And of course check out the paper! www.biorxiv.org/content/10.1...

or @csinva.bsky.social: bsky.app/profile/csin...

I'm posting this thread to highlight some things I thought cool, but if you're interested you should also check out what @rjantonello.bsky.social wrote: bsky.app/profile/rjan...

Cortical weight maps were also reasonably correlated between ECoG and fMRI data, at least for the dimensions well-captured in the ECoG coverage.

Finally, we tested whether the same interpretable embeddings could also be used to model ECoG data from Nima Mesgarani's lab. Despite the fact that our features are less well-localized in time than LLM embeddings, this still works quite well!

To validate the maps we get from this model we also compared them to expectations derived from NeuroSynth and results from experiments targeting specific semantic categories, and also looked at inter-subject reliability. All quite successful.

The model and experts were well-aligned, but there were some surprises, like "Does the input include technical or specialized terminology?" (32), which was much more important than expected.

This method lets us quantitatively assess how much variance different theories explain about brain responses to natural language. So to figure out how well this aligns with what scientists think, we polled experts to see which questions/theories they thought would be important.

"Does the input include dialogue?" (27) has high weights in a smattering of small regions in temporal cortex. And "Does the input contain a negation?" (35) has high weights in anterior temporal lobe and a few prefrontal areas. I think there's a lot of drilling-down we can do here.

The fact that each dimension in the embedding thus corresponds to a specific question means that the encoding model weights are interpretable right out-of-the-box. "Does the input describe a visual experience?" has high weight all along the boundary of visual cortex, for example.

But the wilder thing is how we get the embeddings: by just asking LLMs questions. Each theory is cast as a yes/no question. We then have GPT-4 answer each question about each 10-gram in our natural language dataset. We did this for ~600 theories/questions.

And it works REALLY well! Prediction performance for encoding models is on a par with uninterpretable Llama3 embeddings! Even with just 35 dimensions!!! I find this fairly wild.

So if you're interested in human memory for complex stimuli, please check the paper out! www.biorxiv.org/content/10.1...

This work was a really fun departure for me. Nothing data-driven (and no fMRI!), we just sat down and devised a theory, then tested it. It feels surprisingly good :D

Our model also has interesting linguistic consequences. Speech tends to have uniform information density over time, but there are local variations. We argue that these variations (at least around event boundaries) are actually in service of more uniform _memory_ density.

We also devised a new way to model and study gist with LLMs, which is to measure (or manipulate) the entropy of attention weights for specific "induction" heads within the LLM. Higher entropy attention weights more evenly sample information from the input, and lead to gist-like behavior.