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...

— Dan Goodman (@neural-reckoning.org) 2025-09-18T16:29:52.336Z

Why did we look at what neuromodulation could do in rapid sensory processing? Well, partly just because we could, but also because some evidence suggests there might be something interesting going on here: www.sciencedirect.com/science/arti...

— Dan Goodman (@neural-reckoning.org) 2025-09-18T16:29:52.337Z

We wanted to see what strategies a network would learn with the simplest, most abstract model of neuromodulation, without constraints. So our model is just that a 'neuromodulator' neuron can modify any parameter of another neuron in the same way it modifies a membrane potential.

— Dan Goodman (@neural-reckoning.org) 2025-09-18T16:29:52.338Z

We were also interested in how different temporal and spatial scales affected our results. We model this either by allowing neuromodulator neurons to only act every K time steps, or by forcing them to target a group of G neurons, for various values of K and G.

— Dan Goodman (@neural-reckoning.org) 2025-09-18T16:29:52.339Z

We tested this on three tasks, recognising spoken digits (SHD), commands (SSC) and visual gestures (DVS). In each task, we find much better performance with neuromodulation, tested at network sizes where perforance without neuromodulation had saturated. Each task has a preferred time scale.

— Dan Goodman (@neural-reckoning.org) 2025-09-18T16:29:52.340Z

To our surprise, the spatial scale of neuromodulation didn't have much effect at all but the largest scale. Possibly our tasks just don't capture this well.

— Dan Goodman (@neural-reckoning.org) 2025-09-18T16:29:52.341Z

We wanted to test this in a harder environment, so we looked at a simple cocktail party effect model where speech is presented in an amplitude modulated background noise. This is interesting because this task is hard for state-of-the-art speech recognition systems. Neuromodulation helps a lot!

— Dan Goodman (@neural-reckoning.org) 2025-09-18T16:29:52.342Z

Our next question was to look deeper into the strategy it had learned. Turns out it's doing "listening in the dips", dynamically reducing the number of spikes in noisier parts of the signal by increasing firing thresholds. It does the same thing in natural noise too (see paper).

— Dan Goodman (@neural-reckoning.org) 2025-09-18T16:29:52.343Z

It also seems to improve reaction times, which is kind of cool because this is a role that has previously been suggested for neuromodulation. So far our results on this are a bit preliminary, but we do see a drop in reaction times of around 100ms on a 1s signal.

— Dan Goodman (@neural-reckoning.org) 2025-09-18T16:29:52.344Z

This was a masters project by an exceptional student AbdalQader AlKilany who is starting his PhD with me in a couple of weeks. We'd love to have your feedback on this project so far. It's not submitted for publication yet and we have time to follow up on any ideas. Let us know!

— Dan Goodman (@neural-reckoning.org) 2025-09-18T16:29:52.345Z

Neuromodulation enhances the capability and efficiency of spiking neural networks

AlKilany A, Goodman DFM
Preprint
 

Abstract

Spiking neurons underlie the brain's extreme energy efficiency, and therefore have great potential in neuromorphic computing, although realising this efficiency in practice has proven challenging. We use neuromodulation, a biological mechanism that lets the network dynamically and contextually modify its own parameters. We find it substantially increases performance across a range of sensory processing tasks, including a challenging new speech-in-noise dataset we introduce, with very few additional resources (neurons, energy, parameters). Neuromodulatory networks are space and energy efficient thanks to two mechanisms that are directly relevant to neuromorphic computing and biology: firstly, they allow for an order-of-magnitude reduction in the number of neurons required; and secondly, they enable very sparse firing, achieving better results using orders of magnitude fewer spikes. Together, these properties may throw light on the computational role of neuromodulation in biology, and make neuromodulation an ideal mechanism to improve performance and efficiency for neuromorphic devices.

Links

Categories

The short version

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...

— Dan Goodman (@neural-reckoning.org) 2025-09-18T16:29:52.336Z

Why did we look at what neuromodulation could do in rapid sensory processing? Well, partly just because we could, but also because some evidence suggests there might be something interesting going on here: www.sciencedirect.com/science/arti...

— Dan Goodman (@neural-reckoning.org) 2025-09-18T16:29:52.337Z

We wanted to see what strategies a network would learn with the simplest, most abstract model of neuromodulation, without constraints. So our model is just that a 'neuromodulator' neuron can modify any parameter of another neuron in the same way it modifies a membrane potential.

— Dan Goodman (@neural-reckoning.org) 2025-09-18T16:29:52.338Z

We were also interested in how different temporal and spatial scales affected our results. We model this either by allowing neuromodulator neurons to only act every K time steps, or by forcing them to target a group of G neurons, for various values of K and G.

— Dan Goodman (@neural-reckoning.org) 2025-09-18T16:29:52.339Z

We tested this on three tasks, recognising spoken digits (SHD), commands (SSC) and visual gestures (DVS). In each task, we find much better performance with neuromodulation, tested at network sizes where perforance without neuromodulation had saturated. Each task has a preferred time scale.

— Dan Goodman (@neural-reckoning.org) 2025-09-18T16:29:52.340Z

To our surprise, the spatial scale of neuromodulation didn't have much effect at all but the largest scale. Possibly our tasks just don't capture this well.

— Dan Goodman (@neural-reckoning.org) 2025-09-18T16:29:52.341Z

We wanted to test this in a harder environment, so we looked at a simple cocktail party effect model where speech is presented in an amplitude modulated background noise. This is interesting because this task is hard for state-of-the-art speech recognition systems. Neuromodulation helps a lot!

— Dan Goodman (@neural-reckoning.org) 2025-09-18T16:29:52.342Z

Our next question was to look deeper into the strategy it had learned. Turns out it's doing "listening in the dips", dynamically reducing the number of spikes in noisier parts of the signal by increasing firing thresholds. It does the same thing in natural noise too (see paper).

— Dan Goodman (@neural-reckoning.org) 2025-09-18T16:29:52.343Z

It also seems to improve reaction times, which is kind of cool because this is a role that has previously been suggested for neuromodulation. So far our results on this are a bit preliminary, but we do see a drop in reaction times of around 100ms on a 1s signal.

— Dan Goodman (@neural-reckoning.org) 2025-09-18T16:29:52.344Z

This was a masters project by an exceptional student AbdalQader AlKilany who is starting his PhD with me in a couple of weeks. We'd love to have your feedback on this project so far. It's not submitted for publication yet and we have time to follow up on any ideas. Let us know!

— Dan Goodman (@neural-reckoning.org) 2025-09-18T16:29:52.345Z