Marcus Ghosh

A diagram showing a maze with a gradient (of sensory cues) overlaid.

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

A diagram showing how different architectures (circles) learn distinct input-sensitivities and memory dynamics.

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

A diagram comparing the sample efficiency (learning speed) of two architectures (shown in grey and blue), across 4 maze tasks.

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

A diagram comparing the fitness (task performance) of all pRNN architectures to the fully recurrent architecture, across 4 types of maze environments.

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

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

A diagram showing 2D mazes with gradients of sensory cues.

To compare pRNN function, we introduce a set of multisensory navigation tasks we call *multimodal mazes*.

In these tasks, we simulate networks as agents with noisy sensors, which provide local clues about the shortest path through each maze.

We add complexity by removing cues or walls.

🧵4/9

A diagram showing a feedforward network, a fully recurrent network and 3 of the 126, partially recurrent, architectures between these two extremes.

This allows us to interpolate between:

Feedforward - with no additional pathways.
Fully recurrent - with all nine pathways.

We term the 126 architectures between these two extremes *partially recurrent neural networks* (pRNNs), as signal propagation can be bidirectional, yet sparse.

🧵3/9

A neural network model with: input, hidden and output nodes, and 9 weight matrices.

We start from an artificial neural network with 3 sets of units and 9 possible weight matrices (or pathways).

By keeping the two feedforward pathways (W_ih, W_ho) and adding the other 7 in any combination,

we can generate 2^7 distinct architectures.

All 128 are shown in the post above.

🧵2/9

A diagram showing 128 neural network architectures.

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

Preprint update: The new version of #SPARKS🎇 is out!
Everything's in here: sparks.crick.ac.uk
A thread on what changed 🧵👇
@flor-iacaruso.bsky.social @sdrsd.bsky.social @alexegeaweiss.bsky.social
#neuroskyence #NeuroAI #ML #BioInspiredAI

Happy to discuss what you disagree with?

Description Please note that job descriptions are not exhaustive, and you may be asked to take on additional duties that align with the key responsibilities ment...

Hiring a post-doc at Imperial in EEE. Broad in scope + flexible on topics: neural networks & new AI accelerators from a HW/SW co-design perspective!

w/ @neuralreckoning.bsky.social @achterbrain.bsky.social in Intelligent Systems and Networks group.

Plz share! 🚀: www.imperial.ac.uk/jobs/search-...

Aw thanks!

As I mentioned in the talk, this is my favourite figure!

Though, sadly, it has been confined to the supplement of the upcoming paper.

10. Keep your AIm in mind (🎯)

As a scientist your focus should be on generating insights and understanding, not models with an extra percentage or two in accuracy!

With that in mind, less performant, but more interpretable models may be preferable.

9. Aim for an interpretable, trustworthy model (🤖)

By using methods from explainable AI, we can try to understand why models may make specific predictions.

This can improve trust. Though, remains an open research problem!

8. Add what you know into your model (🦾)

While many AI methods learn from scratch, incorporating prior knowledge (such as physical laws or symmetries) can help and there are a range of techniques for doing this!

7. Start with synthetic data (🧮)

If your model is not working your problem could be:
* Your model (your code + hyperparameters)
* Your data

To resolve this, generate some simple data (e.g. Gaussian points), if your model can't handle data like these, it probably won't work on real data!

6. Start small and simple (🐣)

Many problems don't require a complex model.

Try:
* Establishing a baseline - e.g. guessing randomly or always guessing the mean
* Simple methods - e.g. linear regression
* Then, if necessary, increasingly complex models

FAIR Principles - GO FAIR In 2016, the ‘FAIR Guiding Principles for scientific data management and stewardship’ were published in Scientific Data. The authors intended to provide guidelines to improve the Findability, Accessib...

5. Be FAIR (🤝)

Make sure your work is openly available and easy to reuse.

For example:
* Use @github.com to share code
* @conda.org to share the packages your code needs
* Zenodo to share data and models

www.go-fair.org/fair-princip...

The Good Research Code Handbook This handbook is for grad students, postdocs and PIs who do a lot of programming as part of their research. It will teach you, in a practical manner, how to organize your code so that it is easy to...

4. Invest time in your code (🦄)

It will improve the quality and reproducibility of your work and save you time in the long run!

We recommend following @patrickmineault.bsky.social's excellent Good Research Code Handbook:
goodresearch.dev/index.html

3. Don't reinvent the wheel (🛞)

Most code you need already exists!

Use standard packages (e.g. @scikit-learn.org and @pytorch.org) as much as possible.

And if you are short on data or compute, consider building on existing (pre-trained) models (e.g. @hf.co).

2. Learn some terminology (🗣️)

At first many terms, in papers, talks etc, will seem opaque and confusing.

Getting familiar with these will help your understanding!

We provide a glossary of terms for reference, but really the best way is to read, listen and join seminars or reading groups.

1. Frame your scientific question (🖼️)

Before diving into research, you need to consider your aim and any data you may have.

This will help you to focus on relevant methods and consider if AI methods will be helpful at all.

@scikit-learn.org provide a great map along these lines!

How can we best use AI in science?

Myself and 9 other research fellows from @imperial-ix.bsky.social use AI methods in domains from plant biology (🌱) to neuroscience (🧠) and particle physics (🎇).

Together we suggest 10 simple rules @plos.org 🧵

doi.org/10.1371/jour...

Beyond Rate Coding: Surrogate Gradients Enable Spike Timing Learning in Spiking Neural Networks We investigate the extent to which Spiking Neural Networks (SNNs) trained with Surrogate Gradient Descent (Surrogate GD), with and without delay learning, can learn from precise spike timing beyond fi...

New preprint for #neuromorphic and #SpikingNeuralNetwork folk (with @pengfei-sun.bsky.social).

arxiv.org/abs/2507.16043

Surrogate gradients are popular for training SNNs, but some worry whether they really learn complex temporal spike codes. TLDR: we tested this, and yes they can! 🧵👇

🤖🧠🧪

Had a great time discussing multisensory integration @imrf.bsky.social!

And really enjoyed sharing our new work too

Neuroscience for machine learners A freely available short course on neuroscience for people with a machine learning background. Designed by Dan Goodman and Marcus Ghosh.

@neuralreckoning.bsky.social & I have an online course (videos, text, code)! Neuroscience for those with a machine learning background

neuro4ml.github.io

Modelling Audio-Visual Reaction Time with Recurrent Mean-Field Networks Understanding how the brain integrates multisensory information during detection and decision-making remains an active area of research. While many inferences have been drawn about behavioural outcome...

It’s been a minute (11 years) since my last @imrf.bsky.social
I’m excited to seeing all the great research.
And I’m delighted to give a talk on Friday about Rebecca Brady’s PhD new modelling studies in collab w @bizleylab.bsky.social and @jennycampos.bsky.social
1/2
www.biorxiv.org/content/10.1...

Off to my first @imrf.bsky.social conference!

I'll be giving a talk on Friday (talk session 9) on multisensory network architectures - new work from me & @neuralreckoning.bsky.social.

But say hello or DM me before then!

Had a great time teaching on the @trendcamina.bsky.social summer school in Zambia!

Awesome students and super TAs:
@akoumoundourou.bsky.social, @tomnotgeorge.bsky.social, @jsoldadomagraner.bsky.social, @ashvparker.bsky.social, @pollytur.bsky.social, @skuechenhoff.bsky.social