Marcus Ghosh
Computational neuroscientist.
Research Fellow @imperialcollegeldn.bsky.social and @imperial-ix.bsky.social
Funded by @schmidtsciences.bsky.social
- How does the structure of a neural circuit shape its function? @neuralreckoning.bsky.social & I explore this in our new preprint: doi.org/10.1101/2025... 🤖🧠🧪 🧵1/9
- [Not loaded yet]
- Currently assembling a larger fleet 🚀
- Amid the rise of billion-parameter models, I argue that toy models, with just a few neurons, remain essential—and may be all neuroscience needs, writes @marcusghosh.bsky.social. #neuroskyence www.thetransmitter.org/theoretical-...
- [Not loaded yet]
- That’s exactly the right question! What is the smallest / simplest model for your phenomenon of interest. Some recent work, cited in the article, shows that very small models can be surprisingly powerful. And often we may be far above the lower bound, which then hinders interpretability
- Thanks for reading the article! 1. The multi-billion parameter model discussed in the article was developed by Meta but then fine-tuned by neuroscientists (for research). This could become more common, or neuroscientists could start to train their own "foundation" models from scratch.
- 2. Brains are massively complex, but if our goal is to understand them, then building scale models may not be the best approach. In physics, many breakthroughs have come from abstracting away complexity, as this article highlights! doi.org/10.1111/ejn....
- Toy models, just in time for Christmas! Excited to share my first article for @thetransmitter.bsky.social #neuroskyence
- Amid the rise of billion-parameter models, I argue that toy models, with just a few neurons, remain essential—and may be all neuroscience needs, writes @marcusghosh.bsky.social. #neuroskyence www.thetransmitter.org/theoretical-...
- I'm excited to be teaching with @trendcamina.bsky.social again this summer. Come along!
- Applications Are Open! 🥳 #TReNDCaMinA Summer School 2026 | 29 Jun–15 Jul Dedan Kimathi University of Technology, Kenya For applicants in African countries with backgrounds in neuroscience, medicine, Computer Science, engineering, & related fields. Apply 👉 trendinafrica.org/trend-camina/
- Adam was a truly inspiring scientist. He made neuroscience fun and exciting, and made everything seem possible. The Behaviour and Neural Systems course was formative for many of us. The photo below is from my time @champalimaudr.bsky.social in 2016.
- Nature Sci Rep publishes incoherent AI slop. eLife publishes a paper which the reviewers didn't agree with, making all the comments and responses public with thoughtful commentary. One of these journals got delisted by Web of Science for quality concerns from not doing peer review. Guess which one?
- [Not loaded yet]
- Actually, according to an eLife information session, eLife desk reject ~90% of submissions. So the ones which do get reviewed tend to be high-quality! Even if, like the paper @neural-reckoning.org is referring to, the reviewers don't agree with everything.
- 📍Excited to share that our paper was selected as a Spotlight at #NeurIPS2025! arxiv.org/pdf/2410.03972 It started from a question I kept running into: When do RNNs trained on the same task converge/diverge in their solutions? 🧵⬇️
- Excited to read! You may be interested in our recent work too! We define 128 RNN architectures then compare their behaviour and dynamics (using gradient-based metrics). bsky.app/profile/marc...
- How does the structure of a neural circuit shape its function? @neuralreckoning.bsky.social & I explore this in our new preprint: doi.org/10.1101/2025... 🤖🧠🧪 🧵1/9
- [Not loaded yet]
- [Not loaded yet]
- For the two I've held at Imperial, applications will open summer 2026, for a September 2027 start. * AI in Science (2 years) - www.imperial.ac.uk/ix-ai-in-sci... * ICRF (4 years) - www.imperial.ac.uk/research-and...
- [Not loaded yet]
- I'm always open to talking about postdoctoral fellowships! I'm not quite ready to post my applications online, but am happy to share them privately. Just DM or email me 🙂
- We're almost at the end of the year, and that means an end-of-year review! Send me your favorite NeuroAI papers of the year (preprints or published, late last year is fine too).
- @neural-reckoning.org and I defined a taxonomy of neural network architectures, which we exhaustively explored! We think these models (pRNNs) are exciting for thinking about structure-function relations in artificial and biological neural networks. bsky.app/profile/marc...
- How does the structure of a neural circuit shape its function? @neuralreckoning.bsky.social & I explore this in our new preprint: doi.org/10.1101/2025... 🤖🧠🧪 🧵1/9
- [Not loaded yet]
- [Not loaded yet]
- Congratulations! Hopefully see you somewhere in Europe soon!
- Interested in #neuroscience + #AI and looking for a PhD position? I can support your application @imperialcollegeldn.bsky.social ✅ Check your eligibility (below) ✅ Contact me (DM or email) UK nationals: www.imperial.ac.uk/life-science... Otherwise: www.imperial.ac.uk/study/fees-a...
- Are #NeuroAI and #AINeuro equivalent? @rdgao.bsky.social draws a nice distinction between the two. And introduces Gao's second law: “Any state-of-the-art algorithm for analyzing brain signals is, for some time, how the brain works.” Part 1: www.rdgao.com/blog/2024/01...
- [Not loaded yet]
- For physical systems yes (from hydraulics to computers)! But, I don't remember seeing an algorithm-centric version?
- I've been waiting some years to make this joke and now it’s real: I conned somebody into giving me a faculty job! I’m starting as a W1 Tenure-Track Professor at Goethe University Frankfurt in a week (lol), in the Faculty of CS and Math and I'm recruiting PhD students 🤗
- Congratulations! Exciting to see an inversion of NeuroAI too
- New preprint! What happens if you add neuromodulation to spiking neural networks and let them go wild with it? TLDR: it can improve performance especially in challenging sensory processing tasks. Explainer thread below. 🤖🧠🧪 www.biorxiv.org/content/10.1...
- Really neat! It could be interesting to explore heterogenous neuromodulation? In this case, having multiple modulator networks which act at different timescales and exert different effects on the downstream SNN? This would be a bit closer to neuromodulators in vivo.
- [Not loaded yet]
- Congratulations!
- The misleading manifold? The current debate (decoding vs causal relevance) and a toy example I gave in the thread below got me thinking about a related issue: how decoding may reflect structure more than function. 🧵 1/5
- [Not loaded yet]
- What I was trying to highlight is that it is possible to observe low dimensional structure in population activity which is unrelated to the computation / task. And it's not obvious that this problem vanishes in more complex circuits?
- [Not loaded yet]
- I think by adding a mixture of noise, delays etc, we may not end up at d=1. But even if we did, many studies would describe these "population dynamics" as a line attractor. Rather than just a single neuron acting on it's own?
- What I was trying to highlight is that it is possible to observe low dimensional structure in population activity which is unrelated to the computation / task. And it's not obvious that this problem vanishes in more complex circuits?
- Its good to see that there are causal studies. But, these seem to be the exception? And perhaps the language used in the non-causal studies can be a bit misleading? For instance, the post that launched this thread claimed that "the brain uses distributed coding" (based on solely on recordings).
- [Not loaded yet]
- The toy circuit is very simple. But I'd be happy to say it computes? For example, given LIF dynamics it could perform temporal coincidence detection (only outputting spikes in response to inputs close together in time).
- [Not loaded yet]
- I think by adding a mixture of noise, delays etc, we may not end up at d=1. But even if we did, many studies would describe these "population dynamics" as a line attractor. Rather than just a single neuron acting on it's own.
- However, In this case, from the anatomy (circuit diagram above), we know that only neuron_1 is involved in the computation (transforming the input to the output). And the manifold we observe is misleading. 🧵4/5
- Causal experiments could help to untangle this. But, it seems like this remains underexplored? Keen to hear your thoughts @mattperich.bsky.social, @juangallego.bsky.social and others! www.nature.com/articles/s41... 🧵5/5
- If we start from this circuit: Input -> neuron_1 -> output ↓ ⋮ ↓ neuron_n And record neurons 1 to n simultaneously (where n could be very large). We can obtain a matrix of neural activity (neurons x time). 🧵2/5
- A common approach to analysing this data would be to apply PCA (or another technique). Yielding a matrix of population activity (d x time). Where d < the number of neurons. A common interpretation of this would be that "the brain uses a low dimensional manifold to link this input-output". 🧵3/5
- [Not loaded yet]
- [Not loaded yet]
- Okay, let's say we make the circuit above more realistic: add neurons, separate them into areas, consider multiple tasks etc. The problem outlined above still persists? If a method doesn't work in a simple system (e.g. the one above), there is no reason to think it will in a more complex one?
- [Not loaded yet]
- But what does the lag tell you about the system? Taking the circuit above as an example.
- Input -> neuron_1 -> output ↓ neuron_2 In this "circuit": * We could decode the input from either neuron * But the circuit is not "using" neuron_2 in the computation (transforming the input to output). * And ablations would make this clear? 🧵 (2/3)
- There may be a difference in the lag between the input signal and neurons 1 and 2. But this difference could be too small to be detectable? And without causal experiments the conclusion could be something like "different neurons represent the stimulus at different timescales"? 🧵 (3/3)
- [Not loaded yet]
- I agree that perturbations have their challenges (see doi.org/10.1371/jour... from @kayson.bsky.social for a better approach). But without them you just have correlations? Just to give a toy example: 🧵 (1/3)
- Being part of this grassroots 🌱 neuroscience collaboration was a great experience! Keep an eye out for our next collaborative effort
- These, and other, studies show that you can decode task-related signals from many brain areas. But wouldn't we need causal manipulations to conclude that the brain "uses" them? For example, maybe we can decode equally well from two areas. But, only one impacts behaviour when inactivated.
- I'll be presenting this work at #CCN2025 tomorrow (A173). Come and say hi or message me if you'd like to meet up!
- How does the structure of a neural circuit shape its function? @neuralreckoning.bsky.social & I explore this in our new preprint: doi.org/10.1101/2025... 🤖🧠🧪 🧵1/9
- 100% We study networks with 8 hidden units and between 104 and 442 parameters (depending on the pRNN architecture). So our networks are very small! But not quite as tiny as @marcelomattar.bsky.social - whose RNNs have < 100 parameters.
- Though, training all 128 pRNN architectures on tasks with experimental data, as in the tiny RNN paper, would be really interesting! Then, we could ask questions like: ⭐ Which architectures can best learn those tasks? ⭐ What pathways do they have? ⭐ Which best match experimental data? Etc
- [Not loaded yet]
- For the mazes we just: ⭐ Create maze structures (walls and paths) using a classic maze generation algorithm. ⭐ Add sensory cues - positive gradients leading towards the goal. Then, we add complexity by removing: ⭐ Walls - to create loops. ⭐ Cues - to create gaps in the gradient.
- Good question! For a TLDR - our title: "Partial recurrence enables robust and efficient computation" Tries to capture the key ideas, so: ⭐ Smaller networks (pRNNs) ⭐ Can be robust (learn good solutions which are robust to perturbations) ⭐ Yet, efficient (use less parameters, energy and space)
- Though for a tattoo Fig 6A (all 128 architectures) or Fig 2A (the pRNN diagram) Would probably look better.
- [Not loaded yet]
- Agreed! As @neuralreckoning.bsky.social mentioned, in some preliminary work, we found that our approach and a Shapley estimate gave similar results. But, it would be worth revisiting with our full data set.
- [Not loaded yet]
- Thanks! We focussed on pairwise differences for interpretability, but you're right it would be interesting to look at n-wise changes too. And actually, even exploring up to the maximum difference does seem feasible! Something to include in the revised version.
- Third, to explore why different circuits function differently, we measured 3 traits from every network. We find that different architectures learn distinct sensitivities and memory dynamics which shape their function. E.g. we can predict a network’s robustness to noise from its memory. 🧵8/9
- We’re excited about this work as it: ⭐ Explores a fundamental question: how does structure sculpt function in artificial and biological networks? ⭐ Provides new models (pRNNs), tasks (Multimodal mazes) and tools, in a pip-installable package: github.com/ghoshm/Multi... 🧵9/9
- First, across tasks and functional metrics, many pRNN architectures perform as well as the fully recurrent architecture. Despite having less pathways and as few as ¼ the number of parameters. This shows that pRNNs are efficient, yet performant. 🧵6/9
- Second, to isolate how each pathway changes network function, we compare pairs of circuits which differ by one pathway. Across pairs, we find that pathways have context dependent effects. E.g. here hidden-hidden connections decrease learning speed in one task but accelerate it in another. 🧵7/9
- [Not loaded yet]
- We trained over 25,000 pRNNs on these tasks. And measured their: 📈 Fitness (task performance) 💹 Learning speed 📉 Robustness to various perturbations (e.g. increasing sensor noise) From these data, we reach three main conclusions. 🧵5/9
