Thanks for this. I've been out of the game for a long time now and it's nice to read non-jargony discussion of the issues.
"Statements like these exasperate me—how self-important and entitled are some of these AI researchers, to claim the right to make decisions for so many people in the rest of the world?"
That was a really interesting read! Thank you for writing!
I’m a roboticist who has wandered a little bit away from my usual home, but I can deeply relate to the ever growing overlap and encroaching of AI on your field. Though I’d wish AI people would stop trying to rename our field of robotics as “Physical AI” without our permission.
I have a question that might be silly, so forgive me if it is. But I was really interested in this part of your article,
“That is, there seem to be many parallels between representation learning tasks in artificial neural networks and the brain. There are somewhat fewer in behavioral tasks, or at least, the results are less cleanly matched.”
As a roboticist I’m interested in intelligent systems, not of the super-mind type, but of the low level controlling a body type. What are some resources I could look to to understand this above quote? Are you saying that humans and animals don’t necessarily learn a model representation of ourselves and the environment? Or that we more just think of objective functions and move our body to maximize reward?
I’ve been getting really curious into cogsci and neurosci motor learning lately. I noticed in the past there was some work that crossed between robotics and motor learning via optimal control. Now I see some with RL. But those papers always felt like it was applying OC or RL to try to blanket explain why we move our bodies. Are there works that more try to explain how we move than why? Like a from first principles of motor learning rather than “here’s a theory that might explain it abstractly, but not concretely?”
Thank you! I’m looking forward to reading more of your work and personally trying to be a bit of a dilettante in neurosci/cogsci!
I think the concept of objective functions is not as useful in neuroscience as it has been in robotics; animals change their objectives all the time, it seems to me, and that’s what makes them interesting.
Instead, what makes more sense is the emergence of what looks like goal-directed behavior driven by local interactions. You might be familiar with Auke Ijspeert’s work as a roboticist, and I’ve been really interested in his version of robotic movement—tightly coupled feedback loops can drive all sorts of movement in real bodies, which is super cool. It feels much closer to small-organism movement than something like RL, or at least it may be a necessary complement to the RL perspective.
At the same time, if you have more references on the subject please send them along!
This is a good reminder that neuroscience and AI are now talking to each other, but not always cleanly. Similar tools do not guarantee that brains and large models work in the same way. Do you think AI will clarify neuroscience, or pull it toward misleading analogies?
Re: the IBL mouse datasets, curious if you have read Learning, fast and slow (https://arxiv.org/pdf/2205.02075) and have a take on it?
These bits near the end summarize its view on very slowly-learned 2AFC tasks:
> Given the enormous difference in learning rates between easy and hard tasks (Figure 1) one is forced to ask whether the learning of hard tasks actually plays any role in the life of an animal. For a mouse, the 6 weeks of training required to distinguish two abstract stimuli represents a good fraction of the animal’s life span. Over that period a wild mouse develops from a newborn to having its own pups. It seems unlikely that the wild mouse uses this time to practice any particular action 10,000 times. Any information in the environment that is relevant to the mouse’s fate must be picked up on the first attempt or at most after a few exposures.
...
> However, a small short-lived animal like the mouse may not be the best model system, given that it cannot possibly benefit from ultra-slow learning. Of course, six weeks of training on a repetitive task will take its toll, and engrave some kind of circuit in the plastic brain matter of the mouse. But those artificial circuits will have little in common with what the animal uses naturally for learning and decision-making; nor do they have to align with the mechanisms of slow learning in humans.
I’ve read the book and I COMPLETELY agree with it—it’s actually what Xiaoqin Wang talked about at Cosyne too (I mention him at the end) and what I wanted to write the opinion piece about. Also in line with the octopus playing piano video you posted.
I think it’s not just the length of time, either; animals doing these lab tasks tend to be water- or food-deprived and it just seems like a narrow subset of behaviors we’re really interested in.
But labs like his are looking into other kinds of behaviors to study, e.g. curiosity-driven ones. I’m personally hoping more groups get on that trend
Yeah his lab has a couple interesting papers following up on this thought where they tracked mice's labyrinth solving abilities (https://pmc.ncbi.nlm.nih.gov/articles/PMC8294850/) & then proposed a biologically plausible algorithm that could match them (https://elifesciences.org/reviewed-preprints/84141). The algorithm explains the patrolling/exploration behavior in an interesting way too, using the concept of "neglect". Though it's unclear to me how their algorithm might generalize to non-discrete/labyrinth contexts (eg open fields).
What I found especially strong here is the sense that the field is no longer asking only whether a model performs well, but at what level of description its behavior becomes intelligible.
From a neuroscience perspective, that distinction matters a great deal. Predictive success alone does not yet tell us whether we understand a system in a scientifically useful way. The harder question is whether we can identify the right intermediate language, rich enough to capture real internal structure, but disciplined enough not to collapse into metaphor.
That is what makes your framing so compelling. The most interesting shift may be not simply from brains to models, or from data to scale, but from output-based evaluation toward the search for a workable psychology of the machine.
Thanks for this. I've been out of the game for a long time now and it's nice to read non-jargony discussion of the issues.
"Statements like these exasperate me—how self-important and entitled are some of these AI researchers, to claim the right to make decisions for so many people in the rest of the world?"
More on that 'rest of the world' here.
https://benzhouworld.substack.com/p/the-discipline-that-forgot-why-it
That was a really interesting read! Thank you for writing!
I’m a roboticist who has wandered a little bit away from my usual home, but I can deeply relate to the ever growing overlap and encroaching of AI on your field. Though I’d wish AI people would stop trying to rename our field of robotics as “Physical AI” without our permission.
I have a question that might be silly, so forgive me if it is. But I was really interested in this part of your article,
“That is, there seem to be many parallels between representation learning tasks in artificial neural networks and the brain. There are somewhat fewer in behavioral tasks, or at least, the results are less cleanly matched.”
As a roboticist I’m interested in intelligent systems, not of the super-mind type, but of the low level controlling a body type. What are some resources I could look to to understand this above quote? Are you saying that humans and animals don’t necessarily learn a model representation of ourselves and the environment? Or that we more just think of objective functions and move our body to maximize reward?
I’ve been getting really curious into cogsci and neurosci motor learning lately. I noticed in the past there was some work that crossed between robotics and motor learning via optimal control. Now I see some with RL. But those papers always felt like it was applying OC or RL to try to blanket explain why we move our bodies. Are there works that more try to explain how we move than why? Like a from first principles of motor learning rather than “here’s a theory that might explain it abstractly, but not concretely?”
Thank you! I’m looking forward to reading more of your work and personally trying to be a bit of a dilettante in neurosci/cogsci!
Hi, thanks for reading!
I think the concept of objective functions is not as useful in neuroscience as it has been in robotics; animals change their objectives all the time, it seems to me, and that’s what makes them interesting.
Instead, what makes more sense is the emergence of what looks like goal-directed behavior driven by local interactions. You might be familiar with Auke Ijspeert’s work as a roboticist, and I’ve been really interested in his version of robotic movement—tightly coupled feedback loops can drive all sorts of movement in real bodies, which is super cool. It feels much closer to small-organism movement than something like RL, or at least it may be a necessary complement to the RL perspective.
At the same time, if you have more references on the subject please send them along!
This is a good reminder that neuroscience and AI are now talking to each other, but not always cleanly. Similar tools do not guarantee that brains and large models work in the same way. Do you think AI will clarify neuroscience, or pull it toward misleading analogies?
Thanks for the writeup!
Re: the IBL mouse datasets, curious if you have read Learning, fast and slow (https://arxiv.org/pdf/2205.02075) and have a take on it?
These bits near the end summarize its view on very slowly-learned 2AFC tasks:
> Given the enormous difference in learning rates between easy and hard tasks (Figure 1) one is forced to ask whether the learning of hard tasks actually plays any role in the life of an animal. For a mouse, the 6 weeks of training required to distinguish two abstract stimuli represents a good fraction of the animal’s life span. Over that period a wild mouse develops from a newborn to having its own pups. It seems unlikely that the wild mouse uses this time to practice any particular action 10,000 times. Any information in the environment that is relevant to the mouse’s fate must be picked up on the first attempt or at most after a few exposures.
...
> However, a small short-lived animal like the mouse may not be the best model system, given that it cannot possibly benefit from ultra-slow learning. Of course, six weeks of training on a repetitive task will take its toll, and engrave some kind of circuit in the plastic brain matter of the mouse. But those artificial circuits will have little in common with what the animal uses naturally for learning and decision-making; nor do they have to align with the mechanisms of slow learning in humans.
I’ve read the book and I COMPLETELY agree with it—it’s actually what Xiaoqin Wang talked about at Cosyne too (I mention him at the end) and what I wanted to write the opinion piece about. Also in line with the octopus playing piano video you posted.
I think it’s not just the length of time, either; animals doing these lab tasks tend to be water- or food-deprived and it just seems like a narrow subset of behaviors we’re really interested in.
But labs like his are looking into other kinds of behaviors to study, e.g. curiosity-driven ones. I’m personally hoping more groups get on that trend
Yeah his lab has a couple interesting papers following up on this thought where they tracked mice's labyrinth solving abilities (https://pmc.ncbi.nlm.nih.gov/articles/PMC8294850/) & then proposed a biologically plausible algorithm that could match them (https://elifesciences.org/reviewed-preprints/84141). The algorithm explains the patrolling/exploration behavior in an interesting way too, using the concept of "neglect". Though it's unclear to me how their algorithm might generalize to non-discrete/labyrinth contexts (eg open fields).
What I found especially strong here is the sense that the field is no longer asking only whether a model performs well, but at what level of description its behavior becomes intelligible.
From a neuroscience perspective, that distinction matters a great deal. Predictive success alone does not yet tell us whether we understand a system in a scientifically useful way. The harder question is whether we can identify the right intermediate language, rich enough to capture real internal structure, but disciplined enough not to collapse into metaphor.
That is what makes your framing so compelling. The most interesting shift may be not simply from brains to models, or from data to scale, but from output-based evaluation toward the search for a workable psychology of the machine.