Reposted by Hessam AkhlaghpourSam Gershman @gershbrain.bsky.social · 04/04/2025Bridging computation and representation in associative learning: gershmanlab.com/pubs/Gershma...gershmanlab.com 04512
Hessam Akhlaghpour @thehessam.bsky.social · 02/04/2025Just out by Ellen Lau: How single-neuron computation matters for cognitive neuroscience doi.org/10.21428/8e6...doi.orgHow single-neuron computation matters for cognitive neuroscience 081
Reposted by Hessam AkhlaghpourArula Ratnakar @arula-ratnakar.bsky.social · 28/03/2025Future Tense Fiction Magazine asked me to write a short scifi story for them, which would be paired with a response from a scientist. Inspired by work from Cristopher Moore, @thehessam.bsky.social & @lenoreblum.bsky.social I wrote "Coda" And Cristopher Moore wrote the response essay! Links below: 2218
Reposted by Hessam AkhlaghpourEvelina Tutucci @anilevetu.bsky.social · 16/03/2025From the Besse's lab. #RNAbiology #RNAlocalization #spatialbiology www.nature.com/articles/s41...nature.comAxonal RNA localization is essential for long-term memory - Nature CommunicationsAxonal transcriptomes have been characterized, but their regulation and role in neuronal circuits remain unclear. Here authors uncover activity-dependent RNA localization in axons important for memory... 04514
Reposted by Hessam AkhlaghpourJason Shepherd @jasonsynaptic.bsky.social · 17/02/2025Feels weird to promote science right now but here’s a new review/opinion piece from the lab where we outline new insights about how Arc mediates long lasting synaptic plasticity and memory formation. 🧪🧠 kwnsfk27.r.eu-west-1.awstrack.me/L0/https:%2F...kwnsfk27.r.eu-west-1.awstrack.me 912616
Hessam Akhlaghpour @thehessam.bsky.social · 16/01/2025To add to all the other relevant papers mentioned in the comments: Gaby Maimon, my current advisor, from his grad school years www.cell.com/neuron/fullt...cell.comBeyond Poisson: Increased Spike-Time Regularity across Primate Parietal CortexCortical areas differ in their patterns of connectivity, cellular composition, and functional architecture. Spike trains, on the other hand, are commonly assumed to follow similarly irregular dynamics... 020
Reposted by Hessam AkhlaghpourIpshita Zutshi @ipshitaz.bsky.social · 09/01/2025It’s an incredible feeling to see something you’ve poured your heart into come to life. This work was born from my curiosity about hippocampal “cognitive tuning.” Our interpretation of the environment depends on our actions, constantly evolving as we pursue our goals www.nature.com/articles/s41...nature.comHippocampal neuronal activity is aligned with action plans - NatureUsing high-density electrophysiological recordings, how internally generated cell assemblies are updated by action plans to meet external goals is explored. 518347
Hessam Akhlaghpour @thehessam.bsky.social · 09/01/2025As far as I’m aware it’s still an unproven conjecture 020
Reposted by Hessam AkhlaghpourPhilipp Brand @pbrand.bsky.social · 08/01/2025I'm thrilled to be opening my own lab this Fall at the @maxplanck.de MPI for Brain Research in Frankfurt. The lab is going to investigate the neuronal and genetic underpinnings of behavioral evolution. We are hiring on all levels. Don't hesitate to share widely and reach out if interested. 2114150
Hessam Akhlaghpour @thehessam.bsky.social · 08/12/2024Exactly. Same with TMs and memory. At any given moment it uses a finite amount of memory. Any computation requires finite amount of resources to complete. So where is the difference in terms of being realistic? 200
Hessam Akhlaghpour @thehessam.bsky.social · 08/12/2024Would you say the same about unbounded time and energy? Because RNNs and finite state machines are unbounded in time and energy. Are they unrealistic idealizations too? 110
Hessam Akhlaghpour @thehessam.bsky.social · 08/12/2024Had you said memory expansion in RNNs is expansion in precision, that would be a totally valid response. Unbounded precision is fine, just like unbounded tape length is fine. But the problem with those models is something else - not the unboundedness of memory. Does that make sense? 2/2 100
Hessam Akhlaghpour @thehessam.bsky.social · 08/12/2024The only way I can understand “RNN with unbounded memory” is as “RNN with unbounded precision”. I cannot see how an RNN can use unbounded number of neurons, because # neurons is intrinsically bounded as part of the system’s dynamics (just like FSA is intrinsically memory bounded). 1/2 110
Hessam Akhlaghpour @thehessam.bsky.social · 08/12/2024There is this paper that discusses one part of your question. doi.org/10.1016/j.nl... But I really don’t know. 100 years ago, when people were proposing genetic information is stored as atomic arrangement in molecules, they had no idea about readout (polymerases, ribosomes, etc).doi.orgRedirecting 010
Hessam Akhlaghpour @thehessam.bsky.social · 08/12/2024Yes, that is still the subject. It sounds like you are claiming that RNNs are universal because you can build an RNN that simulates the physical memory-bounded digital computer (say the one on my desk). You can use the same exact argument to say finite state automata are universal, which is wrong. 100
Hessam Akhlaghpour @thehessam.bsky.social · 07/12/2024(Relatedly, Dijkstra once said "Computer science is no more about computers than astronomy is about telescopes.") RNNs are supposed to be an abstract system of computation. Each RNN solves a certain input output function. The comparison should be to other systems of computation. 2/2 100
Hessam Akhlaghpour @thehessam.bsky.social · 07/12/2024I think the last statement is misleading. It depends on how you define "digital computer". The way you are treating it here is as a device that runs programs, not as an abstract system for computation. In that case it is meaningless to assign a scope of solvable functions to that device. 1/3 100
Hessam Akhlaghpour @thehessam.bsky.social · 07/12/2024the state of an RNN cannot be described independent of the number of neurons. In fact, if you assume bounded precision, you can describe the RNNs progression in time as a string of fixed (unchanging) length. This is not the case for, say, combinatory logic or any other universal system. 3/3 120
Hessam Akhlaghpour @thehessam.bsky.social · 07/12/2024The description of a program (or the description of its state at any time during its progression) can be specified as an arbitrary length string. One would need to add memory *during* computation (as part of the rules of the system) in order to execute it properly. In contrast, 2/3 110
Hessam Akhlaghpour @thehessam.bsky.social · 07/12/2024In the same sense that you can "expand" memory of a finite state automata by adding more states? There is a fundamental difference between that and memory expansion in computer programs. The difference is that a progam's state can be described independent of the available memory. 1/3 120
Hessam Akhlaghpour @thehessam.bsky.social · 07/12/2024How are you imagining memory expansion in RNNs? Through increasing precision (like the 1990s models that suffer from structural stability)? Or by increasing number or neurons? If you can clarify this, maybe I’ll be able to convey my point better. 130
Hessam Akhlaghpour @thehessam.bsky.social · 03/12/2024To elaborate with an example: If Moore's 1998 conjecture is wrong and physically realizable RNNs are universal, then RNNs would be a useful system for organisms to have and tweak/learn/evolve upon. That doesn't mean a specifically universal RNN (capable of running any TM) would be advantageous. 3/3 120
Hessam Akhlaghpour @thehessam.bsky.social · 03/12/2024What kinds of problems are in the reach of biological organisms? What kind of computational problems are solvable through evolution, or learning, or some other biological process? Having a language/system with the scope of *all solvable problems* would be adaptive / advantageous. 2/3 120
Hessam Akhlaghpour @thehessam.bsky.social · 03/12/2024Thanks for engaging with these ideas and sharing it here! One point (potential misunderstanding) is about point 2 that you raised. I'm not saying that an individual organism would necessarily benefit from the capability of universal computation. My argument is about scope, not instances. 1/3 120
Reposted by Hessam AkhlaghpourWallace Marshall @wallaceucsf.bsky.social · 02/12/2024Where does pattern come from inside cells? in this revised preprint, @aralbright.bsky.social Shows that regionalization of mRNA in a Stentor cell depends on dynein motors and the microtubule cytoskeleton doi.org/10.1101/2023... 315242
Reposted by Hessam AkhlaghpourDavid Colaço @davidcolaco.bsky.social · 02/12/2024@jonathannajenson.bsky.social and I will discuss @thehessam.bsky.social's work in the next session of our International Memory Reading Group 183
Reposted by Hessam AkhlaghpourGualtiero Piccinini @gualtiero.bsky.social · 02/12/2024@thehessam.bsky.social speculates that universal computation implemented by RNA is behind memory and cognition. It's a cool idea, though probably wrong. 1/ 1112
Reposted by Hessam AkhlaghpourNicole Rust @nicolecrust.bsky.social · 02/12/2024I admire trainees stepping up like this to talk about the big questions. @thehessam.bsky.social does a terrific job here talking about the nature of memory storage (synaptic vs not). 4266
Hessam Akhlaghpour @thehessam.bsky.social · 27/11/2024I almost forgot how in the old twitter days, we could have productive exchanges in the comments. It’s happening again for me on bluesky - where I walk away from a discussion with learning something new to look up! 060
Hessam Akhlaghpour @thehessam.bsky.social · 27/11/2024Yes, I agree - sort of. It's a question of implementability or applicability to physically realizable systems. Wouldn't you agree with Moore's paragraph that I quoted? 110
Hessam Akhlaghpour @thehessam.bsky.social · 27/11/2024Yes. And the problem with those models is not with those assumptions (arbitrary precision is fine, just like arbitrary tape length is fine) but rather structural stability. See my comment here that highlights excepts from Moore 1998 (after Sontag & Signalman papers came out) bsky.app/profile/theh... 240
Hessam Akhlaghpour @thehessam.bsky.social · 27/11/2024My point is not about whether universal computation is implemented in a particular instance of computation. It's simply this: what are the scope of functions you cover in a system of computation? Does it cover all computable functions? People typically say the set of RNNs do. And that's misleading. 140
Hessam Akhlaghpour @thehessam.bsky.social · 27/11/2024My argument mirrors Cris Moore's argument from this 1998 paper: sites.santafe.edu/~moore/pubs/... See screenshots. 051
Reposted by Hessam AkhlaghpourPessoaBrain @pessoabrain.bsky.social · 27/11/2024𝗪𝗵𝗮𝘁 𝗶𝗻 𝘁𝗵𝗲 𝗯𝗿𝗮𝗶𝗻 𝗶𝗺𝗽𝗹𝗲𝗺𝗲𝗻𝘁𝘀 𝘂𝗻𝗶𝘃𝗲𝗿𝘀𝗮𝗹 𝗰𝗼𝗺𝗽𝘂𝘁𝗮𝘁𝗶𝗼𝗻? Maybe not neurons, perhaps RNA?? Great episode of Brain Inspired with Hessam Akhlaghpour. Do dynamical systems and/or RNNs implement universal computation? This discussion alone would be worth having. #neuroscience braininspired.co/podcast/199/braininspired.coBI 199 Hessam Akhlaghpour: Natural Universal Computation | Brain Inspired 5358
Hessam Akhlaghpour @thehessam.bsky.social · 26/11/2024Since my early postdoc years, whenever my hands were busy with experiments but my mind was free, I’d listen to Brain Inspired. I would have been thrilled to know that I’d be on Paul’s show in the future, discussing what I think is the most exciting problem in biology. braininspired.co/podcast/199/braininspired.coBI 199 Hessam Akhlaghpour: Natural Universal Computation | Brain Inspired 0174
Reposted by Hessam AkhlaghpourThe Transmitter @thetransmitter.bsky.social · 26/11/2024Could the brain’s computational abilities extend beyond neural networks to molecular mechanisms? Listen to Paul Middlebrooks's latest "Brain Inspired" podcast, in which he talks with @thehessam.bsky.social about how universal computation may have evolved via RNA mechanisms: bit.ly/4eQ4Idcthetransmitter.orgHessameddin Akhlaghpour outlines how RNA may implement universal computationCould the brain’s computational abilities extend beyond neural networks to molecular mechanisms? Akhlaghpour describes how natural universal computation may have evolved via RNA mechanisms. 0359
Hessam Akhlaghpour @thehessam.bsky.social · 15/11/2024Ooh! “We demonstrate that PSL is Turing Universal”. Reading for the weekend! 010
Reposted by Hessam AkhlaghpourMelanie Mitchell @melaniemitchell.bsky.social · 21/10/2024New post on my Substack: aiguide.substack.com/p/the-llm-re...aiguide.substack.comThe LLM Reasoning Debate Heats UpThree recent papers examine the robustness of reasoning and problem-solving in large language models 3323
Hessam Akhlaghpour @thehessam.bsky.social · 13/10/2024I hadn't! Thanks for mentioning! The first one doesn't present a new model but surveys existing models, all of which are covered in my list. The other says transformers are computationally limited when "the number of computation steps is bounded by the input length n" which is a different question 010
Hessam Akhlaghpour @thehessam.bsky.social · 13/10/2024Thanks! Yes, you made me realize I shouldn’t have used that term here. It is unnecessarily confusing. 010
Hessam Akhlaghpour @thehessam.bsky.social · 13/10/2024What some of these papers have done is to change the definition of a transformer such that it loses this property. They either 1) don’t use tokens from a vocabulary or 2) use hardmax instead of softmax and rely of the tie case of hardmax. Then they show it is Turing-complete. 2/2 100
Hessam Akhlaghpour @thehessam.bsky.social · 13/10/2024Ah, I think I see what you’re saying now. I’m talking about the function that describes the single pass (which is the function that is learned). That function in any [standard] transformer (no matter how its parameters were found, actually) is resilient to some small amount of noise. 1/2 100
Hessam Akhlaghpour @thehessam.bsky.social · 13/10/2024 Take any working LLM (whose parameters came from deep learning and uses softmax attention layers. If you add epsilon to all parameters, nothing will change (for some small enough eps). 100
Hessam Akhlaghpour @thehessam.bsky.social · 13/10/2024A piecewise continuous function is discontinuous, yet contains neighborhoods of similarly behaving parameters almost everywhere (structurally stable parameter configurations). Structural stability is a much weaker condition than continuity or even differentiability. 100