One of the best ways to ensure your science is solid is to ground it to the physical world. Build something that works. Most of #neuroscience isn’t tied to the physical world. The big exception is Brain-Computer Interfaces, and I’m betting the future of neuroscience will advance through #BCI. - ThreadSky

shahabbakht.bsky.social • 96 days ago

One of the best ways to ensure your science is solid is to ground it to the physical world. Build something that works.

Most of #neuroscience isn’t tied to the physical world. The big exception is Brain-Computer Interfaces, and I’m betting the future of neuroscience will advance through #BCI.

Comments

tonyzador.bsky.social•96 days ago

A good fraction of neuroscientists are working on disease. For them, "building something that works" would mean a new treatment (or diagnostic)

Are you saying that that's not "grounded in the physical world"?

lune-bellec.bsky.social•96 days ago

It has been an attractive framework. But (1) it is super hard; (2) the main limiting factor at the moment in many conditions seem to be dataset size; (3) many clinical diagnoses are just constructs that do not correspond to a disease. Using them as targets for prediction is muddy water.

tonyzador.bsky.social•95 days ago

I think that might be more true of psychiatric disease than say neurodegenerative disorders like ALS no? But yes it's hard.

Also, not sure if dataset size rather than say mechanistic insight is what is limiting for eg neurodegeneration

lune-bellec.bsky.social•95 days ago

Also agree that mechanistic insights seem to be lacking -- particularly obvious in neurodevelopmental conditions. I was more thinking of predictive modelling, where there is a more or less clear target (e.g. accuracy of automated diagnostic), and where sample size is an important limiting factor

lune-bellec.bsky.social•95 days ago

ALS seems like a rather well defined clinical entity indeed - although I'm not familiar with it. Even Alzheimer's is partly a moving target. Psychiatry is another affair entirely, and some patients will strongly push back on a "disease" label for diagnosis - e.g. https://journals.sagepub.com/doi/full/10.1177/1362361318820762

shahabbakht.bsky.social•95 days ago

What Lune said, and just to add to that: I agree in general that disease treatment can be a way of anchoring to the physical world, but it depends (1) how connected the symptoms are to the physical world, and (2) how well the symptoms do actually define the disease (related to Lune’s last point).

lune-bellec.bsky.social•96 days ago

I also feel I've been missing a direct feedback loop on my work for too long. Human neuroscience can sometimes feel more like art than science, where people subjectively appreciate the "novelty" and "attractiveness" of a framework, without solid objective measures to ground this assessment on.

miror-ergo-sum.bsky.social•95 days ago

'appreciate the "novelty" and "attractiveness" of a framework, without solid objective measures'
When framed like that, it sounds more like a form of philosophy of mind rather than science. Don't data & discovery fix this though? The unkown being a fertile ground for wild (and beautiful) theorizing?

lune-bellec.bsky.social•95 days ago

Alas in my little corner of neuroimaging research, there's no gold standard for much of anything. And no pervasive culture of fair comparisons and benchmark challenges. That leaves us sometimes in what feels like philosophy or interpretive dance, I'm afraid.

shahabbakht.bsky.social•95 days ago

I’d say it’s not just attractiveness; it’s attractiveness shared among a group of people that also engage in *constructive* criticism based on accepted norms and frameworks. It’s a social construct.

This excellent book is quite relevant here (h/t @dlevenstein.bsky.social)

shahabbakht.bsky.social•95 days ago

Yeah, and it works as long as there is social agreement about the attractiveness and novelty within the subdomain. I think this also underlies why bridging between domains is so challenging, as the subdomains don’t necessarily share each others’ notions of attractiveness.

neurrito.bsky.social•95 days ago

Agreed that anchoring science in physical results is crucial, but I think it’s a stretch to say most neuroscience isn’t tied to the physical. Tons of work on disease models as stated elsewhere, but also basic biology of neurons, synapses, glia, anatomical work, etc etc.

shahabbakht.bsky.social•95 days ago

Of course the brain itself is part of the physical world, so it's, by definition, anchored to it, but we’re not fully leveraging this connection for uncovering its functional principles. Our *models* of the brain should also interact with the physical world in a non-recognizable way.

mathiasgehrig.bsky.social•94 days ago

I still think it's not unlikely that parts of machine learning research drift towards neuromorphic computing which could also help with making the connection eventually. Given the current progress it could take a long time though

nbara.bsky.social•95 days ago

Agree 100%. Felt it even more when I moved to industry. In academia you might get away with showing your BCI provides a ´significant’ bump in accuracy over existing methods. Then move on. But a product that works on average is not enough. Makes you ask questions that are often ignored in literature

shahabbakht.bsky.social•95 days ago

Good point.

emmz.bsky.social•95 days ago

the science problem to engineering problem pipeline :D

shahabbakht.bsky.social•95 days ago

Very close but there is a subtle difference.
I’m not talking about turning the science problem into an engineering one, but using the physical world as a validation ground. If a BCI helps a paraplegic patient walk, it validates our science of neural encoding/decoding in motor control.

shahabbakht.bsky.social•95 days ago

Making it efficient and better would be the engineering pipeline.

shahabbakht.bsky.social•96 days ago

I’d switch to BCI if there weren’t already so brilliant minds working on it.

shahabbakht.bsky.social•96 days ago

#NeuroAI, particularly the pathway from neuroscience to AI, might offer a weak way to anchor neuroscience to the physical world, especially through #robotics. But the gap between the two fields is so deep that the anchor might be too loose to depend on.

andrewtanyongyi.bsky.social•95 days ago

The best is when the thing works and proves the science wrong: STN DBS 🤯🤩

shahabbakht.bsky.social•95 days ago

💯

forsch2024.bsky.social•94 days ago

Fully agree.
And this (the missing connection to physical reality) appears to be the case for a significant number of science & research projects.
Many of these projects would greatly benefit from a detailed and clear answer to a simple question:
What's the ultimate purpose of my project?

mikexcohen.bsky.social•92 days ago

Interesting take, but that would rule out most of the abstract&theoretical mathematics and sciences, without which we wouldn't have all the technology we now take for granted. Also, most neuroscience is grounded by empirical data, which comes from the physical world.

jacobrobinson.bsky.social•95 days ago

Completely agree. I also think there is an opportunity for BCI companies to advance neuroscience because there are questions that are best answered with the scale of data created when technologies are scaled commercially.

shahabbakht.bsky.social•95 days ago

Exactly.
I think what’s currently missing in most BCI companies is interest in basic neuroscience with no obvious near-term financial return. Imo, a DeepMind-like model would revolutionize both neuroscience and BCI industry.

Comments

Posting Rules

Reply