UML co-creator Grady Booch disputes neural networks as biological brain models

VIEWS106REPLIES2

1864TECH

Completely different issue: the original post spoke of this being a side effect of evolution which it most assuredly is not for neural networks. And respectfully, I find your list of references to be at best shallow. Furthermore, none of them are predictive as far as I know: what have we learned from building neural networks that has suggested things that we did not yet know about brains?

AN

Aran Nayebi@aran_nayebi

Actually @Grady_Booch, the convergence btwn deep neural networks & brains isn't a loose analogy -- in fact, they're the best *quantitative* predictors of large-scale brain data we have today. Our theory explains this happens when networks are minimal solutions to hard tasks. As a result, not all networks match the brain equally well, and has told us quite powerful things about the brain over the past decade, across brain areas & species. Here's an incomplete sample I've compiled of what we learn about the brain from this AI-driven approach, that we didn't (and couldn't!) know before:

BOOKMARKS1

AN Aran Nayebi@aran_nayebi

2h ago · 51 Views

1399TECH

@Grady_Booch @dyamins Nope, quite the same issue in fact. The evolution part refers to biological neural networks. The statement, "none of them are predictive" does not concur with empirical reality from the past decade. And the explanations in the slides already state what we have learned.

GB

Grady Booch@Grady_Booch

Completely different issue: the original post spoke of this being a side effect of evolution which it most assuredly is not for neural networks. And respectfully, I find your list of references to be at best shallow. Furthermore, none of them are predictive as far as I know: what have we learned from building neural networks that has suggested things that we did not yet know about brains?

LIKES2

GB Grady Booch@Grady_Booch

1h ago · 28 Views

1864TECH

@aran_nayebi @dyamins Very good, then. Thank you for your patience with me. Sigh. There is so much to read, and so little time...

AN

Aran Nayebi@aran_nayebi

There's a few studies to point to that support this. Perhaps the closest experiment done in biological brains to what we did in ANNs is Malone et al. 2024 Nature Comms: https://www.nature.com/articles/s41467-024-45853-4 They imaged large MEC populations while mice learned a novel virtual environment over ten days. Position decoding from just 50 randomly selected MEC cells (their Fig. 4g) improved with learning & correlated with behavioral performance, and they showed that the consistency of non-grid cells was higher than that of grid cells. Most importantly, optogenetically disrupting the spatial consistency of the population impaired memory-guided navigation. I'd love to see more experiments in this direction where there are further cell-type-specific refined knockouts -- it's often difficult to optogenetically target *functional* cell types (and the advantage of models is we can do this far more quickly than the biologists can, allowing us to generate predictions!)

GB Grady Booch@Grady_Booch

3h ago · 73 Views

1864TECH

Let me give you an example to make it clear what I’m trying to say. Consider the paper you referenced by Nayebi. They first studied organic medial entorhinal cortex cells. The rest of the paper basically talks about how they learned from data that to improve neural models of the brain region. Pretty much the rest of their paper talks about in silico simulations of those whose result improved existing neural networks but – like good science should do - I can see no findings that suggested deeper understanding of the organic mechanisms, predictions from those models that might be observable in your organic system, or for that matter any suggestions to guide further data collection. In short, in that paper, I found nothing that suggests they discovered things about the brain that we did not already know, but rather they found ways to improve their models. Is this useful of course. What does this “teach us about the brain”? That is not evident at all. Ergo my original post stands.

GB

Grady Booch@Grady_Booch

Completely different issue: the original post spoke of this being a side effect of evolution which it most assuredly is not for neural networks. And respectfully, I find your list of references to be at best shallow. Furthermore, none of them are predictive as far as I know: what have we learned from building neural networks that has suggested things that we did not yet know about brains?

GB Grady Booch@Grady_Booch

2h ago · 54 Views

@aran_nayebi @dyamins Then I stand corrected

AN Aran Nayebi@aran_nayebi

2h ago · 32 Views

I am the Nayebi of the paper you reference :) Actually, our knockout experiments in Fig. 4D make the prediction that the border & grid cells people have long overstudied do not drastically matter for navigation compared to the heterogeneous cells that make up most of the region & have long been ignored/naturally arise through optimization of place-cell-based integration.

GB Grady Booch@Grady_Booch

2h ago · 42 Views

@aran_nayebi @dyamins A reasonable follow up: Has that prediction been verified by experiment on organic cells? (I found no repeatable experiments in the literature.) If yes, then I was most definitely wrong. If not, then I stand by my original post.

GB Grady Booch@Grady_Booch

2h ago · 23 Views

1864TECH

@phillip_isola @dyamins Again, a different issue. The original claim was that this was evidence of convergent evolution between organic and artificial neurons.

PI

Phillip Isola@phillip_isola

@Grady_Booch @dyamins Less sure about the neuroscience side, but studying this kind of convergence is absolutely actionable in developing better ML algs. Identifying convergent structures informs us about how to create interfaces between different nets. e.g. https://arxiv.org/abs/2602.17584 and many others

PI Phillip Isola@phillip_isola

2h ago · 8 Views

@Grady_Booch @dyamins I just want to push back on your characterization that this work is simply amusing but not actionable. In ML it already is. In neuro one could imagine implications toward future brain-machine interfaces.

GB Grady Booch@Grady_Booch

1h ago · 19 Views

1864TECH

@phillip_isola @dyamins I can imagine many things. But that is not one of them, or at best, it is one that is very far in the future.

PI

Phillip Isola@phillip_isola

@Grady_Booch @dyamins I just want to push back on your characterization that this work is simply amusing but not actionable. In ML it already is. In neuro one could imagine implications toward future brain-machine interfaces.