@gruntled-neurotech.bsky.social
I see. So Kilosort is on the green end? It models the population of single neurons that lead to recorded electrical activity but you don't really derive scientific insight from its product.
Either way, "strong NeuroAI" feels like a steelman. "AI as models" and "AI as tools" are both strong NeuroAI.
Either way, "strong NeuroAI" feels like a steelman. "AI as models" and "AI as tools" are both strong NeuroAI.
January 19, 2026 at 9:04 PM
I see. So Kilosort is on the green end? It models the population of single neurons that lead to recorded electrical activity but you don't really derive scientific insight from its product.
Either way, "strong NeuroAI" feels like a steelman. "AI as models" and "AI as tools" are both strong NeuroAI.
Either way, "strong NeuroAI" feels like a steelman. "AI as models" and "AI as tools" are both strong NeuroAI.
On the blue end, I can see how useful brain-inspired models of some phenomenon (say, image recognition) could turn out not to be great models of neural computations or physiology. But on the green end, what good are the models of the brain if they are not brain-like?
January 18, 2026 at 9:15 PM
On the blue end, I can see how useful brain-inspired models of some phenomenon (say, image recognition) could turn out not to be great models of neural computations or physiology. But on the green end, what good are the models of the brain if they are not brain-like?
NIH study sections seem to be... whimsical... about what is and isn't 'allowed'. Don't put links in it, I guess. I find even the wording of hypotheses in grants variable. Some allude to mechanisms, some are predictions about experiments, some are questions, some are straight-up model descriptions.
January 16, 2026 at 6:02 PM
NIH study sections seem to be... whimsical... about what is and isn't 'allowed'. Don't put links in it, I guess. I find even the wording of hypotheses in grants variable. Some allude to mechanisms, some are predictions about experiments, some are questions, some are straight-up model descriptions.
Isn't at least a part of it about storytelling? In many cases, the figure 7 model could easily be figure 2 with a slightly prophetic narrative.
January 14, 2026 at 6:34 PM
Isn't at least a part of it about storytelling? In many cases, the figure 7 model could easily be figure 2 with a slightly prophetic narrative.
Experimentalists sort of already do this informally. But I 100% agree that we should do it more and do it more formally with theorists. The pushback you mentioned in your original post is a little odd. If 'doing theory' helps my experiments, I'd certainly welcome it.
January 12, 2026 at 6:59 PM
Experimentalists sort of already do this informally. But I 100% agree that we should do it more and do it more formally with theorists. The pushback you mentioned in your original post is a little odd. If 'doing theory' helps my experiments, I'd certainly welcome it.
Hmm. There is a kind of modeling exercise that could theoretically constrain hypotheses and refine experimental parameters before doing the experiment. But after the experiment is done, I don't quite understand to what end modeling should precede analysis? I'm going to do both anyway, right?
January 12, 2026 at 6:25 PM
Hmm. There is a kind of modeling exercise that could theoretically constrain hypotheses and refine experimental parameters before doing the experiment. But after the experiment is done, I don't quite understand to what end modeling should precede analysis? I'm going to do both anyway, right?
This is practically obviously true but the dichotomy between what drives science (hypothesis or discovery) is dissolving rapidly. If you ask anyone why they did an experiment, they will lay out a very clear, even mechanistic, hypothesis. But their analyses are incredibly data- and discovery-driven.
January 12, 2026 at 5:32 PM
This is practically obviously true but the dichotomy between what drives science (hypothesis or discovery) is dissolving rapidly. If you ask anyone why they did an experiment, they will lay out a very clear, even mechanistic, hypothesis. But their analyses are incredibly data- and discovery-driven.
There are models as analytical tools (encoding/decoding models, say), and models that create artificial versions of some phenomena (CNN models of visual computation, say). The latter thrives on data but doesn't need it like the former.
January 12, 2026 at 5:27 PM
There are models as analytical tools (encoding/decoding models, say), and models that create artificial versions of some phenomena (CNN models of visual computation, say). The latter thrives on data but doesn't need it like the former.
I don't know whether this new policy will be transformative. But I also don't think this changes anything for existing Canadian departments already rife with foreign-trained talent. Your point that this money could be better spent to help make existing PIs more productive is well-taken though.
January 8, 2026 at 4:32 PM
I don't know whether this new policy will be transformative. But I also don't think this changes anything for existing Canadian departments already rife with foreign-trained talent. Your point that this money could be better spent to help make existing PIs more productive is well-taken though.
Fair. But to your point about 'undermining training pipeline' and 'you have to leave to be valued', I find it telling that to a first approximation, no PI at, say, McGill Biology, was locally trained. If Canada wanted to recruit locally, they would be doing that already.
January 7, 2026 at 5:54 PM
Fair. But to your point about 'undermining training pipeline' and 'you have to leave to be valued', I find it telling that to a first approximation, no PI at, say, McGill Biology, was locally trained. If Canada wanted to recruit locally, they would be doing that already.
It won't happen quickly but there are only so many students. If there are three new labs in a department of ten, graduate recruitment does not go up 30%. Nevertheless, my point is that I can't see how more labs is a bad thing.
January 7, 2026 at 5:13 PM
It won't happen quickly but there are only so many students. If there are three new labs in a department of ten, graduate recruitment does not go up 30%. Nevertheless, my point is that I can't see how more labs is a bad thing.
Regarding the retaining local trainee point, every neurobiology lab I know overproduces PhDs. Canada doesn't have enough postdoc slots, let alone TT positions. Increasing labs and reducing PhDs is the way to create a self-sustaining research ecosystem. 2/2
January 7, 2026 at 5:02 PM
Regarding the retaining local trainee point, every neurobiology lab I know overproduces PhDs. Canada doesn't have enough postdoc slots, let alone TT positions. Increasing labs and reducing PhDs is the way to create a self-sustaining research ecosystem. 2/2
This assumes a zero-sum game. What do you think happens once you have recruited the 20 international PIs? You have 20 more PIs that the government needs to fund over the next 20-40 years. The pot of money will have to expand naturally. Isn't that a good thing? 1/2
January 7, 2026 at 5:02 PM
This assumes a zero-sum game. What do you think happens once you have recruited the 20 international PIs? You have 20 more PIs that the government needs to fund over the next 20-40 years. The pot of money will have to expand naturally. Isn't that a good thing? 1/2
So how would you interpret fMRI patches not corroborated by electrophysiology and causal perturbations? If you're relying on convergence, there must be some intrinsic explanation for when that doesn't happen. In other words, what good are the fMRI discovered patches?
January 5, 2026 at 3:19 PM
So how would you interpret fMRI patches not corroborated by electrophysiology and causal perturbations? If you're relying on convergence, there must be some intrinsic explanation for when that doesn't happen. In other words, what good are the fMRI discovered patches?
It was script-by-committee. They wanted to give whale goo immortality, human-cordyceps breathing symbiosis, fire tribe supremacy, reanimated human retrieving long-lost son, and Navi girl-Jesus squid-summoner subplots the same, dire stakes. And they ignored the unobtanium plot from the first movie!
January 2, 2026 at 9:38 PM
It was script-by-committee. They wanted to give whale goo immortality, human-cordyceps breathing symbiosis, fire tribe supremacy, reanimated human retrieving long-lost son, and Navi girl-Jesus squid-summoner subplots the same, dire stakes. And they ignored the unobtanium plot from the first movie!
Just because the decision in a laboratory task was seemingly based on an image on the screen, doesn't make it a purely perceptual decision. Also applies for visually-guided movements vs spontaneous movements. Pointless debates in papers and on the internet.
December 14, 2025 at 10:26 PM
Just because the decision in a laboratory task was seemingly based on an image on the screen, doesn't make it a purely perceptual decision. Also applies for visually-guided movements vs spontaneous movements. Pointless debates in papers and on the internet.
It helps if I read this as '_we_ weren't super excited about this at this moment' than a value judgement. Sucks either way.
December 12, 2025 at 7:29 PM
It helps if I read this as '_we_ weren't super excited about this at this moment' than a value judgement. Sucks either way.
These data are from a recognition memory task which we have famously high capacity for. To begin to address how these high/low D representations are used to guide behavior, we can ask, is the spectral dimensionality different when the subjects report correctly or incorrectly? 3/3
December 12, 2025 at 4:58 PM
These data are from a recognition memory task which we have famously high capacity for. To begin to address how these high/low D representations are used to guide behavior, we can ask, is the spectral dimensionality different when the subjects report correctly or incorrectly? 3/3
In papers that report low D, the low variance dims are so because the high variance dimensions explain behavior. This really nice paper affirms previous results -- if behavior is not incredibly constrained by low d aspects of the stimuli/actions, activity is free to vary in many dimensions. 2/3
December 12, 2025 at 4:58 PM
In papers that report low D, the low variance dims are so because the high variance dimensions explain behavior. This really nice paper affirms previous results -- if behavior is not incredibly constrained by low d aspects of the stimuli/actions, activity is free to vary in many dimensions. 2/3
I'm a fan of this paper but I don't think previous studies report low dimensionality because the dataset wasn't large enough and, if I'm understanding the second reason correctly, the low-variance dimensions have never been meaningless. The utility of the dimensions for behavior matters hugely. 1/3
December 12, 2025 at 4:58 PM
I'm a fan of this paper but I don't think previous studies report low dimensionality because the dataset wasn't large enough and, if I'm understanding the second reason correctly, the low-variance dimensions have never been meaningless. The utility of the dimensions for behavior matters hugely. 1/3
Hmm. I’d go further. If the participants were simply rating the same images on a scale of average redness to greenness, you’d see drastically lower dimensionality across cortex. This is an empirical question, of course.
December 12, 2025 at 5:22 AM
Hmm. I’d go further. If the participants were simply rating the same images on a scale of average redness to greenness, you’d see drastically lower dimensionality across cortex. This is an empirical question, of course.
Wait so if the participants in this paper were using the images to do something useful, you would see lower dimensionality?
December 12, 2025 at 4:13 AM
Wait so if the participants in this paper were using the images to do something useful, you would see lower dimensionality?
