One thing AI has clarified is that generalization emerges from recursive compression across many environments, not from solving isolated paradigms cleanly. Toy tasks often capture a slice of behavior but miss the dynamics that determine when and how a system updates its representations. That gap seems central to the limits of many cognitive theories.
I linked to the following in a comment about your Article #2 on the topic, but it's directly relevant to this article: [Sloman-1993-Prospects for AI is the General Science of Intelligence](https://cogaffarchive.org/Aaron.Sloman_prospects.pdf). He argues what the title says. I highly recommend the article because it reframes the design stance in a much more useful way than Daniel Dennett and David Marr did.
I think I disagree with the following passage, or I’m misunderstanding it:
“In the first post of this series, I mentioned that minimalistic task design is often a design principle of cognitive (neuro)science, but one that also might be holding thefield back, by focusing our understanding of systems onto narrow settings. I think that the discussion above offers a different perspective on these issues: if richer and larger-scale settings can fundamentally change what a system learns and how it generalizes, than studying learning & generalization solely in toy settings may mislead us.”
Most cognitive neuroscience research assesses cognitive/brain systems of people who have experienced diverse set of stimuli in a wide variety of contexts (i.e., in the person’s life prior to the experiments). So, it makes sense to study the solution the brain has come up with experiments that test specific hypotheses. For example, if you test a specific hypothesis about vision by manipulating artificial stimuli, and you find that the model of object classification does not behave like a human following the manipulation, it suggests that the model is has found a different solution to classifying objects, no? The specific experimental setup used in CogNeuro is critical for characterising how the mind and the model work – not holding us back.
Of course, it is always possible that a domain specific model developed in CogNeuro only succeeds with artificial stimuli associated with a specific phenomenon, and an ANN does better with naturalistic stimuli (e.g., gets a better Brain-Score). In which case, both models are deficient as models of the brain. But I think there is a misunderstanding of what CogNeuro researchers are trying to do.
For instance, you write:
“Instead, we argue that these simpler components need to be grounded in models that can really perform the same range of naturalistic tasks as the natural intelligences do, in as similar a range of naturalistic settings as possible.”
But most researchers in CogNeuro are *not* trying to build a complete model of the visual system, etc. They are trying to characterise specific properties of the mind/brain, and when modelling these properties, they are trying to gain insights to these specific functions. ANNs that do well on Brain-Score need to behave like humans in these experiments if they are taken to be models of humans.
At the bottom of this post is a list of findings about human vision that any model of the human vision needs to contend with (the passage is taken from our response to the commentaries of our BBS article). You might want to argue that the brain adopted these specific solutions because it has been trained on a wide variety of contexts of naturalistic stimuli, and thus it is necessary to train your model this way. But they way to know whether you have succeeded in building an ANN that sees like humans is to assess whether the model shows these phenomena.
Finally, the problem with many ANNs of humans is that they are trained on orders of magnitude more data than humans. This suggests that researchers need to build much more innate structure in their models. I don’t think pretraining using back-prop is a reasonable proxy for evolution. For example, here is a link to a paper of mine that shows how the massive amount of training of LLMs shows that LLMs are missing innate inductive biases that humans enjoy: https://doi.org/10.1037/rev0000595
From BBS response:
“To give only the most cursory of overviews, the following findings should play a central role in theory and model building. The input to our visual system is degraded due to a large blind spot and an inverted retina with light having to pass through multiple layers of retinal neurons, axons and blood vessels before reaching the photoreceptors. Nevertheless, we are unaware of the degraded signals due to a process of actively filling in missing signals in early visual cortex (e.g., Grossberg, 2003; Ramachandran, & Gregory, 1991). We have fovea that support high-acuity colour vision for only about 2 degrees of visual angle (about the size of a thumbnail at arm’s length). Nevertheless, we have the subjective sense of a rich visual experience across a much wider visual field because we move our eyes approximately 3 times per second (Rayner, 1978), with the encoding of visual inputs suppressed during each saccade (Matin, 1974), and the visual system somehow integrating inputs across fixations (Irwin, 1991). At the same time, we can identify multiple objects in scenes following a single fixation (Biederman, 1972), with object identification taking approximately 150 ms (Thorpe et al., 1996) - too quick to rely on recurrence. We are also blind to major changes in a scene as revealed by change blindness (Simons, & Levin, 1997) and have a visual short-term memory of approximately four items (Cowan, 2001). Our visual system organizes image contours by various Gestalt rules to separate figure from ground (Wagemans et al., 2012) and organize contours to build representations of object parts (Biederman, 1987). Objects are encoded in terms of their surfaces, parts, and relations between parts to build 3D representations relying on monocular and binocular inputs (Biederman, 1987; Marr, 1982; Nakayama, & Shimojo, 1992). Colour, form, and motion processing are factorized to the extent that it is possible to be cortically colour blind (Cavanagh et al., 1998), or suffer motion blindness where objects disappear during motion but are visible and recognizable while static (Zeki, 1991), or show severe impairments with object identification while maintaining the ability to reach and manipulate objects (Goodale, & Milner, 1992). Participants can even classify objects while denying seeing them (Koculak & Wierzchon). Our visual system manifests a wide range of visual, size, and shape constancies to estimate the distal properties of the world independent of the lighting and object pose, and we suffer from size, colour and motion illusions that reflect the very mechanisms that serve the building of these distal representations from the proximal image projected onto our retinas. These representations of distal stimuli in the world support a range of visual tasks, including object classification, navigation, grasping, and visual reasoning. All this is done with spiking networks composed of neurons with a vast range of morphologies that vary in ways relevant to their function, with architectures constrained by evolution and biophysics.”
NeuroAI is useful to the extent that it provides explanations for these and countless other experimental findings reported in psychology and neuroscience. Otherwise, ANNs should be considered amazing engineering artefacts that work differently than humans.
To be clear, we are quite explicit in the paper that we value explaining function on simple experimental paradigms too, and that they can be good ways to make controlled test settings. I think we generally agree that we'd like models to explain a broad range of phenomena, and that current models have not achieved even predictive success across the full set of tasks we care about. Still, I maintain that these models are qualitatively better than prior models in terms of their generality, and thinking about the tasks that they fail that some other model could solve can allow us to identify what components they might be missing. I think that's one of the exciting things about these models, and it's certainly something that you ought to be able to contribute to if you understand the system so much better — augment one of these models with features that make it better able to explain some empirical phenomenon you care about, without making it incapable of doing anything else. I think we could all learn something from that.
I also think that it's important to solve the problems with the generality that humans solve them — e.g. for a task like same-different, one can hand-craft a model to solve the problem in a certain narrow set of stimuli, but can we really be confident that is how humans solve it across all stimuli? Focusing on too narrow a task space results in aliasing where many solutions can work.
You seem to move back and forth between wanting models to explain all phenomena before you will consider them useful at all, and wanting to argue that they are worse than models that explain only a single phenomenon, and don't incorporate most of the features you're describing as essential. Again, if there's something models are missing, why not show how to fix it?
Regarding data quantity: I think your point is somewhat valid for language models; for vision models it's much less clear that the amount of data you need is out of the realm of possibility. Both would likely benefit from many other features like interactive learning etc. — lots to be explored there!
Thanks for quick response, but I think we disagree on some basic points.
Having a model that can identify images of naturalistic objects, can generate captions, reason, etc. is great when evaluating it as an engineering tool. But it is not a very useful finding when evaluating a model as a model of human intelligence. The basic problem is that there are multiple ways to solve problems, and the critical question for a vision scientist is whether ANNs solve visual (or other) tasks like humans. Success on correlational benchmark tasks (e.g., most tasks on Brain-Score and the like) is not a strong test.
Indeed, we already know that ANNs often succeed in non-human-like ways. For example, Geirhos et al. showed that ANNs often classify objects based on texture, whereas humans largely rely on shape. At the same time, many of the top scoring models on Brain-Score classify objects based on texture. The predictions (correlations) are not a good measure of mechanistic alignment.
More recently, Dujmović et al. (2024) demonstrated how RSA scores can also be misleading, for the same reason – correlations do not imply causality because of confounds. And Malhotra and Bowers (2024) showed the Brain-Scores in V4 and IT are largely the product of the background of images not the objects themselves, undermining the claim that Brain-Scores provide a measure of “core object recognition”. Indeed, there are reports that LLMs can predict brain activations in humans (Doerig et al., 2025), and even low-level vision in macaques (Conwell et al., 2024). Do you think the good predictions of LLMs are providing any insights into how the visual area V1 in macaques works?
If correlational studies do not demonstrate that ANNs work like biological intelligence, what should you do? You should do like all other sciences do – run experiments that manipulate variables to test hypotheses and rule out confounds. That is, it is not just nice to test ANNs on phenomena and experiments, it is the only way to determine whether the ANN is doing something human-like. The ability of ANNs to solve a range of engineering tasks does not provide any evidence they area human-like.
If you are developing a model of human vision, the key question is whether ANNs provide some insights into the dozen or so visual phenomena I listed at the end of my last post. If not, then what have we learned about human vision from ANN?
You write:
“You seem to move back and forth between wanting models to explain all phenomena before you will consider them useful at all, and wanting to argue that they are worse than models that explain only a single phenomenon, and don't incorporate most of the features you're describing as essential. Again, if there's something models are missing, why not show how to fix it?”
I don’t understand this. I’m not claiming that you need a model to explain all phenomena. I’m quite impressed with a model that can do a good job accounting for a range of experimental results in a specific domain. Like the old days where PDP models (and earlier) models of word identification were tested against a wide range of experimental results with words. We recently found that various ANNs could account for a range of word priming phenomena (Yin et al., 2023), and that seems to me better evidence that there is something to be learned about word identification with these models than correlational studies that are typically used.
I am on board asking whether you can fix an ANN that does well on correlational benchmarks. And that involves testing the ANN on a wide range of experimental results that have been reported in literature. A key problem with NeuroAI is that editors and grant panels are not much interested in testing ANNs in ways that show they are not human-like. You need to provide “solutions” – much easier to do when solutions are operationalised as high correlations. Falsification is not much appreciated in NeuroAI. It is also just important to carry out experiments that test out hypotheses about how human intelligence works without running any computational models.
Finaly, you write
“Regarding data quantity: I think your point is somewhat valid for language models; for vision models it's much less clear that the amount of data you need is out of the realm of possibility”.
Maybe, but in Bowers (2025) I show that current claims that ANNs can learn to label objects with a human-diet of data are unwarranted. The same is true for models that claim to learn to classify objects with a human diet of data – models are trained on many epochs (children only live once), and the measures of success are extremely weak, etc.
Bowers, J. S. (2025). The successes and failures of artificial neural networks (ANNs) highlight the importance of innate linguistic priors for human language acquisition. Psychological Review.
Conwell, C., McMahon, E., Jagadeesh, A. V., Vinken, K., Sharma, S., Prince, J. S., ... & Livingstone, M. (2024). Monkey see, model knew: large language models accurately predict human and macaque visual brain activity. In UniReps: 2nd Edition of the Workshop on Unifying Representations in Neural Models.
Doerig, A., Kietzmann, T. C., Allen, E., Wu, Y., Naselaris, T., Kay, K., & Charest, I. (2025). High-level visual representations in the human brain are aligned with large language models. Nature Machine Intelligence, 7(8), 1220-1234.
Dujmović, M, Bowers, J.S., Adolfi, F., & Malhotra, G. (2024). Inferring DNN-Brain alignment using representational similarity analyses can be problematic. International Conference on Learning Representations workshop conference proceeding. https://openreview.net/forum?id=dSEwiAENTS
Yin, D., Biscione, V., & Bowers, J. S. (2023). Convolutional neural networks trained to identify words provide a surprisingly good account of visual form priming effects. Computational Brain & Behavior, 6(3), 457-472.
You seem to be assuming that I advocated for correlational analyses; that isn't the case (and indeed, there's a whole section in this post, and a longer one in the corresponding paper, on "building naturalistic paradigms without giving up experimental control"). I think we agree on the importance of experiments.
However, many studies have gone beyond correlational analyses to explicitly test the model-brain link experimentally, for example by making strong predictions from them, and observing that they hold in experimental verification (e.g. https://www.nature.com/articles/s41562-023-01783-7).
You say that "I’m not claiming that you need a model to explain all phenomena. I’m quite impressed with a model that can do a good job accounting for a range of experimental results in a specific domain." — I think evidence like the above clearly shows that these models can do a good job accounting for neural data *and* making novel predictions within a domain (e.g. processing of short sentences, in the case of that study). I don't see how this case is so different from the model of science you are advocating for.
But I also disagree that accounting for "a range of experimental results in a specific domain" is always sufficient for guaranteeing understanding, and here I think that both classic cognitive models and present deep learning ones are still inadequate. My motivation for this is actually *precisely* the same as yours at some level; even explaining a set of experimental results does not mean the model implements the solution in the same way. As you well know, there are often debates in the literature that get resolved by a new experiment that illustrates a flaw in prior models. And many oversimplified cognitive tasks, in particular, seem to me to admit many solutions that simply couldn't scale to solve the real problems. That's precisely why I believe accounting for a broader sets of experimental data is *important* for being sure the explanation is not overfit to your task paradigm.
In our paper, we discuss two case studies that I find quite informative for this broader theme: this nice paper from Anne Collins that (among other things) shows how you need to push RL task complexity sufficiently high to disentangle different models that could account for behavior: https://www.nature.com/articles/s41562-025-02340-0 and this beautiful series of works by Tyler Bonnen: https://www.cell.com/neuron/fulltext/S0896-6273(21)00459-1 and https://elifesciences.org/articles/84357 that show how you a model that can actually process the naturalistic stimuli used in prior experiments can reconcile seeming inconsistencies in the literature about whether PRC is involved in perception. These studies (and others we review) illustrate how increasing task complexity and naturalism, and accounting for a broader range of experimental paradigms, can be necessary to resolve competing models and achieve deeper understanding.
“…many studies have gone beyond correlational analyses to explicitly test the model-brain link experimentally, for example by making strong predictions from them, and observing that they hold in experimental verification (e.g. https://www.nature.com/articles/s41562-023-01783-7).”
I agree this is an experimental design where the authors manipulated stimuli to test a hypothesis (assessing whether they could predict which sentences drive the language network more or less based on a LLM). But I disagree it is a strong prediction: the authors find that more complex sentences drive more activation in both the LLM and human language network, and sentences that are easy to process drive both less. I expect many qualitatively models would make this prediction, and indeed, I expect many people would predict this outcome without any model. That is, it is not a severe test of the model (Bowers et al., 2023).
You write: “But I also disagree that accounting for "a range of experimental results in a specific domain" is always sufficient for guaranteeing understanding, and here I think that both classic cognitive models and present deep learning ones are still inadequate.”
This is not a fair characterisation of what I wrote. I’m not claiming that this is sufficient for guaranteeing understanding. Yes, I am impressed when a model can account for a wide range of experiments, but of course, more testing is always required, and there are no guarantees in science.
Very briefly, you also cite work by Bonnen et al. regarding whether the PRC is involved in perception. This work is predicated on their claim “a deep learning framework that approximates visual behaviors supported by the VVS (i.e., lacking PRC).”
But I don’t accept the premise. Deep networks do not approximate the behaviour supported by the VVS (e.g., see recent paper in Nature Machine Intelligence that shows that vision- language models fail on most of the psychological experiments in MindSet vision by https://www.nature.com/articles/s42256-026-01179-y), amongst other failures. Indeed, the brain-scores in high level vision are largely driven by the background of the images, not the objects themselves, undermining the claim that brain-scores reflect “core-object recognition” (Malhotra and Bowers, 2024). Given this, I don’t think the authors conclusion that “this work resolves decades of seemingly inconsistent findings surrounding PRC involvement in perception” is warranted.
I’m not against using naturalistic stimuli, nor using image computable models. But there are so many basic findings in psychology that provide key information about how humans identify objects amongst other visual functions. It would seem to me that these findings should be on the top of the list to explain if you want to claim that ANNs identify objects like humans.
Bowers, J. S., Malhotra, G., Adolfi, F., Dujmović, M., Montero, M. L., Biscione, V., ... & Heaton, R. F. (2023c). On the importance of severely testing deep learning models of cognition. Cognitive Systems Research, 82, 101158.
One thing AI has clarified is that generalization emerges from recursive compression across many environments, not from solving isolated paradigms cleanly. Toy tasks often capture a slice of behavior but miss the dynamics that determine when and how a system updates its representations. That gap seems central to the limits of many cognitive theories.
I linked to the following in a comment about your Article #2 on the topic, but it's directly relevant to this article: [Sloman-1993-Prospects for AI is the General Science of Intelligence](https://cogaffarchive.org/Aaron.Sloman_prospects.pdf). He argues what the title says. I highly recommend the article because it reframes the design stance in a much more useful way than Daniel Dennett and David Marr did.
He published another noteworthy article on AI for cognitive science in 2008 in the Journal of Artificial Intelligence: [https://cogaffarchive.org/sloman-aij-08.pdf](https://cogaffarchive.org/sloman-aij-08.pdf). That's in the same issue and in response to John McCarthy's article in the same issue (also worth reading).
Actually Sloman spent much of his career arguing that: [Aaron Sloman - Google Scholar](https://scholar.google.com/citations?hl=en&user=lTc3UwsAAAAJ). He is regarded by some as Britain's top philosopher of AI/ cognitive science (along with Margaret A. Boden and Andy Clark). I studied under all 3 of them while doing my Ph.D. in UK. (Boden was my external examiner). Hence I reference him a lot. See my [Homage to Aaron Sloman, Winner of the 2020 APA K. Jon Barwise Prize](https://cogzest.com/2020/06/homage-to-aaron-sloman-winner-of-the-2020-apa-k-jon-barwise-prize/) if you're curious.
I think I disagree with the following passage, or I’m misunderstanding it:
“In the first post of this series, I mentioned that minimalistic task design is often a design principle of cognitive (neuro)science, but one that also might be holding thefield back, by focusing our understanding of systems onto narrow settings. I think that the discussion above offers a different perspective on these issues: if richer and larger-scale settings can fundamentally change what a system learns and how it generalizes, than studying learning & generalization solely in toy settings may mislead us.”
Most cognitive neuroscience research assesses cognitive/brain systems of people who have experienced diverse set of stimuli in a wide variety of contexts (i.e., in the person’s life prior to the experiments). So, it makes sense to study the solution the brain has come up with experiments that test specific hypotheses. For example, if you test a specific hypothesis about vision by manipulating artificial stimuli, and you find that the model of object classification does not behave like a human following the manipulation, it suggests that the model is has found a different solution to classifying objects, no? The specific experimental setup used in CogNeuro is critical for characterising how the mind and the model work – not holding us back.
Of course, it is always possible that a domain specific model developed in CogNeuro only succeeds with artificial stimuli associated with a specific phenomenon, and an ANN does better with naturalistic stimuli (e.g., gets a better Brain-Score). In which case, both models are deficient as models of the brain. But I think there is a misunderstanding of what CogNeuro researchers are trying to do.
For instance, you write:
“Instead, we argue that these simpler components need to be grounded in models that can really perform the same range of naturalistic tasks as the natural intelligences do, in as similar a range of naturalistic settings as possible.”
But most researchers in CogNeuro are *not* trying to build a complete model of the visual system, etc. They are trying to characterise specific properties of the mind/brain, and when modelling these properties, they are trying to gain insights to these specific functions. ANNs that do well on Brain-Score need to behave like humans in these experiments if they are taken to be models of humans.
At the bottom of this post is a list of findings about human vision that any model of the human vision needs to contend with (the passage is taken from our response to the commentaries of our BBS article). You might want to argue that the brain adopted these specific solutions because it has been trained on a wide variety of contexts of naturalistic stimuli, and thus it is necessary to train your model this way. But they way to know whether you have succeeded in building an ANN that sees like humans is to assess whether the model shows these phenomena.
Finally, the problem with many ANNs of humans is that they are trained on orders of magnitude more data than humans. This suggests that researchers need to build much more innate structure in their models. I don’t think pretraining using back-prop is a reasonable proxy for evolution. For example, here is a link to a paper of mine that shows how the massive amount of training of LLMs shows that LLMs are missing innate inductive biases that humans enjoy: https://doi.org/10.1037/rev0000595
From BBS response:
“To give only the most cursory of overviews, the following findings should play a central role in theory and model building. The input to our visual system is degraded due to a large blind spot and an inverted retina with light having to pass through multiple layers of retinal neurons, axons and blood vessels before reaching the photoreceptors. Nevertheless, we are unaware of the degraded signals due to a process of actively filling in missing signals in early visual cortex (e.g., Grossberg, 2003; Ramachandran, & Gregory, 1991). We have fovea that support high-acuity colour vision for only about 2 degrees of visual angle (about the size of a thumbnail at arm’s length). Nevertheless, we have the subjective sense of a rich visual experience across a much wider visual field because we move our eyes approximately 3 times per second (Rayner, 1978), with the encoding of visual inputs suppressed during each saccade (Matin, 1974), and the visual system somehow integrating inputs across fixations (Irwin, 1991). At the same time, we can identify multiple objects in scenes following a single fixation (Biederman, 1972), with object identification taking approximately 150 ms (Thorpe et al., 1996) - too quick to rely on recurrence. We are also blind to major changes in a scene as revealed by change blindness (Simons, & Levin, 1997) and have a visual short-term memory of approximately four items (Cowan, 2001). Our visual system organizes image contours by various Gestalt rules to separate figure from ground (Wagemans et al., 2012) and organize contours to build representations of object parts (Biederman, 1987). Objects are encoded in terms of their surfaces, parts, and relations between parts to build 3D representations relying on monocular and binocular inputs (Biederman, 1987; Marr, 1982; Nakayama, & Shimojo, 1992). Colour, form, and motion processing are factorized to the extent that it is possible to be cortically colour blind (Cavanagh et al., 1998), or suffer motion blindness where objects disappear during motion but are visible and recognizable while static (Zeki, 1991), or show severe impairments with object identification while maintaining the ability to reach and manipulate objects (Goodale, & Milner, 1992). Participants can even classify objects while denying seeing them (Koculak & Wierzchon). Our visual system manifests a wide range of visual, size, and shape constancies to estimate the distal properties of the world independent of the lighting and object pose, and we suffer from size, colour and motion illusions that reflect the very mechanisms that serve the building of these distal representations from the proximal image projected onto our retinas. These representations of distal stimuli in the world support a range of visual tasks, including object classification, navigation, grasping, and visual reasoning. All this is done with spiking networks composed of neurons with a vast range of morphologies that vary in ways relevant to their function, with architectures constrained by evolution and biophysics.”
NeuroAI is useful to the extent that it provides explanations for these and countless other experimental findings reported in psychology and neuroscience. Otherwise, ANNs should be considered amazing engineering artefacts that work differently than humans.
To be clear, we are quite explicit in the paper that we value explaining function on simple experimental paradigms too, and that they can be good ways to make controlled test settings. I think we generally agree that we'd like models to explain a broad range of phenomena, and that current models have not achieved even predictive success across the full set of tasks we care about. Still, I maintain that these models are qualitatively better than prior models in terms of their generality, and thinking about the tasks that they fail that some other model could solve can allow us to identify what components they might be missing. I think that's one of the exciting things about these models, and it's certainly something that you ought to be able to contribute to if you understand the system so much better — augment one of these models with features that make it better able to explain some empirical phenomenon you care about, without making it incapable of doing anything else. I think we could all learn something from that.
I also think that it's important to solve the problems with the generality that humans solve them — e.g. for a task like same-different, one can hand-craft a model to solve the problem in a certain narrow set of stimuli, but can we really be confident that is how humans solve it across all stimuli? Focusing on too narrow a task space results in aliasing where many solutions can work.
You seem to move back and forth between wanting models to explain all phenomena before you will consider them useful at all, and wanting to argue that they are worse than models that explain only a single phenomenon, and don't incorporate most of the features you're describing as essential. Again, if there's something models are missing, why not show how to fix it?
Regarding data quantity: I think your point is somewhat valid for language models; for vision models it's much less clear that the amount of data you need is out of the realm of possibility. Both would likely benefit from many other features like interactive learning etc. — lots to be explored there!
Thanks for quick response, but I think we disagree on some basic points.
Having a model that can identify images of naturalistic objects, can generate captions, reason, etc. is great when evaluating it as an engineering tool. But it is not a very useful finding when evaluating a model as a model of human intelligence. The basic problem is that there are multiple ways to solve problems, and the critical question for a vision scientist is whether ANNs solve visual (or other) tasks like humans. Success on correlational benchmark tasks (e.g., most tasks on Brain-Score and the like) is not a strong test.
Indeed, we already know that ANNs often succeed in non-human-like ways. For example, Geirhos et al. showed that ANNs often classify objects based on texture, whereas humans largely rely on shape. At the same time, many of the top scoring models on Brain-Score classify objects based on texture. The predictions (correlations) are not a good measure of mechanistic alignment.
More recently, Dujmović et al. (2024) demonstrated how RSA scores can also be misleading, for the same reason – correlations do not imply causality because of confounds. And Malhotra and Bowers (2024) showed the Brain-Scores in V4 and IT are largely the product of the background of images not the objects themselves, undermining the claim that Brain-Scores provide a measure of “core object recognition”. Indeed, there are reports that LLMs can predict brain activations in humans (Doerig et al., 2025), and even low-level vision in macaques (Conwell et al., 2024). Do you think the good predictions of LLMs are providing any insights into how the visual area V1 in macaques works?
If correlational studies do not demonstrate that ANNs work like biological intelligence, what should you do? You should do like all other sciences do – run experiments that manipulate variables to test hypotheses and rule out confounds. That is, it is not just nice to test ANNs on phenomena and experiments, it is the only way to determine whether the ANN is doing something human-like. The ability of ANNs to solve a range of engineering tasks does not provide any evidence they area human-like.
If you are developing a model of human vision, the key question is whether ANNs provide some insights into the dozen or so visual phenomena I listed at the end of my last post. If not, then what have we learned about human vision from ANN?
You write:
“You seem to move back and forth between wanting models to explain all phenomena before you will consider them useful at all, and wanting to argue that they are worse than models that explain only a single phenomenon, and don't incorporate most of the features you're describing as essential. Again, if there's something models are missing, why not show how to fix it?”
I don’t understand this. I’m not claiming that you need a model to explain all phenomena. I’m quite impressed with a model that can do a good job accounting for a range of experimental results in a specific domain. Like the old days where PDP models (and earlier) models of word identification were tested against a wide range of experimental results with words. We recently found that various ANNs could account for a range of word priming phenomena (Yin et al., 2023), and that seems to me better evidence that there is something to be learned about word identification with these models than correlational studies that are typically used.
I am on board asking whether you can fix an ANN that does well on correlational benchmarks. And that involves testing the ANN on a wide range of experimental results that have been reported in literature. A key problem with NeuroAI is that editors and grant panels are not much interested in testing ANNs in ways that show they are not human-like. You need to provide “solutions” – much easier to do when solutions are operationalised as high correlations. Falsification is not much appreciated in NeuroAI. It is also just important to carry out experiments that test out hypotheses about how human intelligence works without running any computational models.
Finaly, you write
“Regarding data quantity: I think your point is somewhat valid for language models; for vision models it's much less clear that the amount of data you need is out of the realm of possibility”.
Maybe, but in Bowers (2025) I show that current claims that ANNs can learn to label objects with a human-diet of data are unwarranted. The same is true for models that claim to learn to classify objects with a human diet of data – models are trained on many epochs (children only live once), and the measures of success are extremely weak, etc.
Bowers, J. S. (2025). The successes and failures of artificial neural networks (ANNs) highlight the importance of innate linguistic priors for human language acquisition. Psychological Review.
Conwell, C., McMahon, E., Jagadeesh, A. V., Vinken, K., Sharma, S., Prince, J. S., ... & Livingstone, M. (2024). Monkey see, model knew: large language models accurately predict human and macaque visual brain activity. In UniReps: 2nd Edition of the Workshop on Unifying Representations in Neural Models.
Doerig, A., Kietzmann, T. C., Allen, E., Wu, Y., Naselaris, T., Kay, K., & Charest, I. (2025). High-level visual representations in the human brain are aligned with large language models. Nature Machine Intelligence, 7(8), 1220-1234.
Dujmović, M, Bowers, J.S., Adolfi, F., & Malhotra, G. (2024). Inferring DNN-Brain alignment using representational similarity analyses can be problematic. International Conference on Learning Representations workshop conference proceeding. https://openreview.net/forum?id=dSEwiAENTS
Malhotra, G., & Bowers, J.S. (2024). Predicting brain activation does not license conclusions regarding DNN-brain alignment: The case of Brain-Score. Cognitive Computational Neuroscience Conference Proceeding, Boston. https://2024.ccneuro.org/pdf/309_Paper_authored_Neural_Predictivity_CCN2024.pdf
Yin, D., Biscione, V., & Bowers, J. S. (2023). Convolutional neural networks trained to identify words provide a surprisingly good account of visual form priming effects. Computational Brain & Behavior, 6(3), 457-472.
You seem to be assuming that I advocated for correlational analyses; that isn't the case (and indeed, there's a whole section in this post, and a longer one in the corresponding paper, on "building naturalistic paradigms without giving up experimental control"). I think we agree on the importance of experiments.
However, many studies have gone beyond correlational analyses to explicitly test the model-brain link experimentally, for example by making strong predictions from them, and observing that they hold in experimental verification (e.g. https://www.nature.com/articles/s41562-023-01783-7).
You say that "I’m not claiming that you need a model to explain all phenomena. I’m quite impressed with a model that can do a good job accounting for a range of experimental results in a specific domain." — I think evidence like the above clearly shows that these models can do a good job accounting for neural data *and* making novel predictions within a domain (e.g. processing of short sentences, in the case of that study). I don't see how this case is so different from the model of science you are advocating for.
But I also disagree that accounting for "a range of experimental results in a specific domain" is always sufficient for guaranteeing understanding, and here I think that both classic cognitive models and present deep learning ones are still inadequate. My motivation for this is actually *precisely* the same as yours at some level; even explaining a set of experimental results does not mean the model implements the solution in the same way. As you well know, there are often debates in the literature that get resolved by a new experiment that illustrates a flaw in prior models. And many oversimplified cognitive tasks, in particular, seem to me to admit many solutions that simply couldn't scale to solve the real problems. That's precisely why I believe accounting for a broader sets of experimental data is *important* for being sure the explanation is not overfit to your task paradigm.
In our paper, we discuss two case studies that I find quite informative for this broader theme: this nice paper from Anne Collins that (among other things) shows how you need to push RL task complexity sufficiently high to disentangle different models that could account for behavior: https://www.nature.com/articles/s41562-025-02340-0 and this beautiful series of works by Tyler Bonnen: https://www.cell.com/neuron/fulltext/S0896-6273(21)00459-1 and https://elifesciences.org/articles/84357 that show how you a model that can actually process the naturalistic stimuli used in prior experiments can reconcile seeming inconsistencies in the literature about whether PRC is involved in perception. These studies (and others we review) illustrate how increasing task complexity and naturalism, and accounting for a broader range of experimental paradigms, can be necessary to resolve competing models and achieve deeper understanding.
Hi Andrew, you write:
“…many studies have gone beyond correlational analyses to explicitly test the model-brain link experimentally, for example by making strong predictions from them, and observing that they hold in experimental verification (e.g. https://www.nature.com/articles/s41562-023-01783-7).”
I agree this is an experimental design where the authors manipulated stimuli to test a hypothesis (assessing whether they could predict which sentences drive the language network more or less based on a LLM). But I disagree it is a strong prediction: the authors find that more complex sentences drive more activation in both the LLM and human language network, and sentences that are easy to process drive both less. I expect many qualitatively models would make this prediction, and indeed, I expect many people would predict this outcome without any model. That is, it is not a severe test of the model (Bowers et al., 2023).
You write: “But I also disagree that accounting for "a range of experimental results in a specific domain" is always sufficient for guaranteeing understanding, and here I think that both classic cognitive models and present deep learning ones are still inadequate.”
This is not a fair characterisation of what I wrote. I’m not claiming that this is sufficient for guaranteeing understanding. Yes, I am impressed when a model can account for a wide range of experiments, but of course, more testing is always required, and there are no guarantees in science.
Very briefly, you also cite work by Bonnen et al. regarding whether the PRC is involved in perception. This work is predicated on their claim “a deep learning framework that approximates visual behaviors supported by the VVS (i.e., lacking PRC).”
But I don’t accept the premise. Deep networks do not approximate the behaviour supported by the VVS (e.g., see recent paper in Nature Machine Intelligence that shows that vision- language models fail on most of the psychological experiments in MindSet vision by https://www.nature.com/articles/s42256-026-01179-y), amongst other failures. Indeed, the brain-scores in high level vision are largely driven by the background of the images, not the objects themselves, undermining the claim that brain-scores reflect “core-object recognition” (Malhotra and Bowers, 2024). Given this, I don’t think the authors conclusion that “this work resolves decades of seemingly inconsistent findings surrounding PRC involvement in perception” is warranted.
I’m not against using naturalistic stimuli, nor using image computable models. But there are so many basic findings in psychology that provide key information about how humans identify objects amongst other visual functions. It would seem to me that these findings should be on the top of the list to explain if you want to claim that ANNs identify objects like humans.
Bowers, J. S., Malhotra, G., Adolfi, F., Dujmović, M., Montero, M. L., Biscione, V., ... & Heaton, R. F. (2023c). On the importance of severely testing deep learning models of cognition. Cognitive Systems Research, 82, 101158.
Malhotra, G., & Bowers, J.S. (2024). Predicting brain activation does not license conclusions regarding DNN-brain alignment: The case of Brain-Score. Cognitive Computational Neuroscience Conference Proceeding, Boston. https://2024.ccneuro.org/pdf/309_Paper_authored_Neural_Predictivity_CCN2024.pdf