Brad Aimone

I was amused to find myself quoted, kind of, in the last paragraph of this; don't know if I've ever been the punchline quote before!

To me it is the only thing that matters. We all understand how basic ANNs work--we always have actually--it just isn't satisfying because they're trivial and boring. Just model / data fitting and scale.

The brain does it better. How? That's the interesting question.

I think this is the big leverage point few acknowledge. Artificial RNNs are unconstrained, kind of by definition. Universal. But the brain has real constraints in what it computes, how fast, robust, small, connected, developed, and so forth.

Those biology details people hate are the constraints

Cue the C Elegans cynics who say we can't even figure out a 302 neuron network

We are trying hard to get Science Bluesky to the network effects of the good old days!

Neuroscientists allowing other fields to define what details are relevant has been a total disaster. No other discipline would outsource a question of such great importance.

Machine learning is kind of a hack here, because it says "we can't actually define what language processing/image classification/etc is, but we can use data to give us that constraint. So let us just approximate this unknown function that we can't define well"

It works, but it is inefficient.

Incidentally, my personal favorite interpretation of advancements in computer science is that most of the progress we've made has been in defining the problems themselves. Many successes correlate to clearer definitions of what "good" means or abstracting the task away from weak definitions.

This is a great thread, and I think it hits on one of the biggest challenges in neuroscience that I hope NeuroAI can impact.

Until we can rigorously define what computations are occurring in the brain, we can't make any real progress in our functional understanding. We need constraints. 🧠🤖🧪

I wouldn't call myself a Markramian but I am most certainly an anti-Marrian. The Marr tradition has allowed cog sci and AI to ignore biology and biologists to ignore computation. So while we know a lot about AI and a lot about the biology, we know little about the Brain. It's been a total disaster.

I don't think it is quite a fair comparison. The vast majority of systems neuroscientists have been banging some version of the Marr drum for 50 years. Maybe a few dozen have really tried the Markram approach?

Take away Markram's personality clashes, I think the jury is out on that approach

I was amazed when I learned a few years ago that many English academics have an "at work accent" and a "relaxed at home around family and local friends accent" which is extremely different and that they can turn them on and off as needed.

Wikipedia, which is probably the best internet consolidation of science for the masses, is heavily biased at the edges towards those self-servicing scientists who want to hype themselves.

So clearly they shouldn't replicate and sell copyrighted material. But if you read my paper and incorporate that into your thinking, that isn't something you need to ask permission

Accessing illegally is a different story. But this isn't an assault on Science; I'd argue it is almost ideal...

To be fair, isn't this why we publish? To get out ideas integrated into the universal human knowledge-base?

It would be far worse if Meta's AI was trained *without* your papers, wouldn't it be?

Neuromorphic Simulation of Drosophila Melanogaster Brain Connectome on Loihi 2 We demonstrate the first-ever nontrivial, biologically realistic connectome simulated on neuromorphic computing hardware. Specifically, we implement the whole-brain connectome of the adult Drosophila ...

We put the FlyWire connectome on the Loihi 2 neuromorphic platform

🤖🧠🧪🪰

arxiv.org/abs/2508.16792

We comp neuros can't make anyone happy can we?

I'm just bitter about all of the AI crowd and their strawman about planes not having feathers.

The modern AI approach would be to put a nuclear fueled engine on the back of that wooden airplane. And then condescendingly mock neuroscientists by saying the wings aren't actually important either.

Yeah, models don't have to be perfect. But details aren't inherently bad, they just may not be relevant to your task and may be needed elsewhere

We see this acutely in AI. Turns out the brain's details weren't that important for image classification. So now we have LLMs that are 1,000,000x too big.

So the value function often becomes something like "simpler is better" or "focus on what you understand" which are orthogonal concepts to "important for the brain"

This is always a dangerous point when it comes to the brain, as this assumes you have a clear and proper objective function

My opinion is that we have little idea what flight even means for neural computation, so we have little basis on determining which details are important.

We aren't that far removed from neuroscientists celebrating the perceived failure of the Human Brain Project and its computational goals.

I'm not sure, there is definitely a similarity between neuromorphic and FPGAs in terms of how algorithms should be represented but I don't know much about the theory around FPGAs

From a serial complexity on a conventional platform I think you're right. For a true event driven data flow architecture like neuromorphic, I don't think there should have to be a program counter or similar (there isn't a program per se). Not sure what is really under the hood on today's platforms

Sitting at the ModSim meeting this week, it is clear that every other scientific field has benefited more from advanced computing than neuroscience. Those of us who have tried to do large scale modeling know who has been dismissive in the past. The lack of serious computing in neuro is neuro's fault

An interesting comment and discussion that gets to the deeper question of whether NeuroAI is a move towards something new or repackaging of old tired approaches with a shiny AI paint job.

Traffic in Seattle is really bad...

I worry that serious neuroscientists avoid "emotion" as a rigorous area of study because it seems too close to the quack scientists who fight about consciousness

Yet emotion is central to the major neural health challenges we face as a society. People aren't on SSRIs because their V2 is misbehaving

You can get really good store guac at a few places in Texas. Sometimes safer than buying avocados which can be random.

I feel this way though about all pasta sauces. A clove of garlic and a can of tomatoes makes a better marinara than any jarred stuff

(incidentally, I think this is the same local attractor we are stuck in in neuroscience, with experimental systems and theories)