I love this project, which feels like it opens many doors.

Great opportunity here!!

My impression is that there is a cultural barrier where systems neuroscientists overlook the relevance of non-invasive brain imaging to test certain hypotheses, simply because it's not directly measuring spikes.

My point is that it's too simplistic to think that measuring spiking activity is always a more "direct" way of testing hypotheses about the relationship between neural activity and information processing.

It's also possible that the patterns are encoded in a distributed way across thousands of neurons. To test that hypothesis, you might need methods like 2 photon imaging.

In some cases, what matters for information processing is inter-area communication, which is not something that is measured in most ephys studies, but can be accessed with fMRI (again, I'm not arguing that's true in the case of imagery).

Ephys will only sample a relatively small number of neurons at the same time. It won't find patterns that are only visible when you aggregate many neurons. It's conceivable that methods like fMRI could detect such patterns (I'm not arguing that's true in the case of imagery).

Let's take the hypothesis that sensory neurons are activated during mental imagery. There are trade-offs in using different methods to address this question. And it depends a lot on how you think about the relevant neural activity.

I think my point here has been misunderstood by most commentators. Obviously I agree that ephys is a more direct measure of spiking activity. Let me try to clarify...

Is the alternative hypothesis that something other than neural activity is generating the non-invasive brain signals? Or is the qualifier "direct" simply a rhetorical flourish?

Papers that report neuronal recordings sometimes say that they provide "direct" evidence for neural encoding of something, in contrast to "indirect" methods such as non-invasive brain imaging. I don't really understand this. What makes one "direct" and the other "indirect"?

Please don't tell me about perceptrons, neocognitrons, or TD learning. These aren't good examples, for reasons I go into here: https://t.co/6M04CYYdfD

I'm mainly interested in examples where neuroscientists make a correct ex ante prediction that some brain fact will improve performance when incorporated into an AI system (without changing other things). Ex post explanations in terms of neuroscience are less convincing.

I'd like to teach a paper which shows how a fact about the brain materially improved an AI system in a way that is unlikely to have been figured out by engineering alone. I haven't been able to find a single example of this. Suggestions welcome.

What are the most compelling real-world applications of cognitive science? I'm looking for empirical papers (not opinions/reviews).

we've finally updated our 2⃣ year old preprint on action chunking as (conditional) policy compression! this was a lesson in perseverance and not giving up on your ideas...work with @gershbrain & @AnnHuang42 ✨ (1/3)

I advised a physics student and I kept worrying that the physics police would burst in to take her away because she wasn’t doing “real” physics. But I realized that if it’s interesting to physicists, it’s physics.

[Please retweet!] My paper, 'Cognitive Representations of Social Relationships and their Developmental Origins" has been accepted at Behavioral and Brain Sciences! I'm thrilled to be able to engage with people's commentary! Please consider writing one!

If you're interested in the future of scholarly publishing, check out this panel discussion (open to the public): https://t.co/umpgjsGN6w

(5') If you are invited to review and you don't have time, please just decline! Don't leave the invitation hanging, because this just slows down the whole process.