About
What do I do?
Most neural networks compute by pushing neuron activations forward through a network, layer by layer. The neural networks I work on compute by falling downhill — every neuron settling into a local minimum of an overarching energy landscape at the same time. That difference matters because a physical system can do the settling for you: it is the principle behind neuromorphic hardware, brain-inspired chips that compute with physics instead of simulating it on a GPU, for a fraction of the energy.
- Algorithms: I develop learning rules that train these systems without backpropagation.
- Hardware Applications: I show how the algorithms map onto real physical devices, where the settling is done by physics rather than being simulated.
- Landscapes: I study what determines whether a system settles somewhere useful or whether it gets stuck.
Why do I do it?
Two reasons. The first is energy. The human brain runs on about 20 watts (about as much as a lightbulb 💡), while training a large model in a datacentre takes thousands of times more, and the model still can’t match the generality of the brain’s intelligence. New ‘neuromorphic’ hardware that computes by settling, inspired by models of the human brain, could close much of that gap.
The second is that relaxation on an energy landscape is one of the oldest ideas about how brains might learn (see Hopfield networks and predictive coding networks etc). The machine learning community largely set it aside, not because it was necessarily wrong but because it ran slowly on GPUs. Enabling these models to run quickly on neuromorphic hardware lets us take the idea seriously again, at modern scale, and explore potential advantages of these approaches.
Ultimately we are both wasting electricity, and missing something crucial about what makes human intelligence so efficient. By understanding this better, we can help protect the planet and uncover more about human intelligence at the same time.
What do I really do?
The longer version
Everything I currently work on relates to energy landscapes: systems whose dynamics are (noisy) gradient descent to a local minimum, and the algorithms that train them. That covers energy-based and predictive coding models in machine learning; learning directly on neuromorphic hardware, where the settling is performed by physics rather than simulated; and the statistical physics of slow relaxation, which asks what properties of landscapes lets those dynamics reach a good minimum instead of getting trapped in a bad one.
My PhD, joint between the Math & Algorithms Group at Nokia Bell Labs and the Theoretical Condensed Matter Group at the University of Cambridge, centres on Equilibrium Propagation (EP), which computes loss gradients from a system’s own equilibria instead of a backward pass.
I first showed that Oscillator Ising Machines — CMOS oscillator networks originally built for combinatorial optimisation — have energy gradient descent dynamics and therefore can be trained using EP to state-of-the-art neuromorphic accuracy, and that training survives realistic hardware constraints. A hardware implementation is currently in its early stages with Nokia and TU Eindhoven.
I then developed GradEP, which extends EP from training the locations of local energy minima to training the gradient of the energy landscape. That makes flow matching, score matching and energy-based generation trainable with local rules and therefore on neuromorphic hardware. Its first application, FlowEqProp, was the first flow-matching generative model trained by EP, and won Best Paper at ICONS ‘26.
Most recently, in PotentialEP, I turned to predictive-coding networks, hierarchical energy models first proposed as a theory of the cortex. I proved that adding a single linear term to their energy makes the relaxed network output the gradient of an Energy Matching potential. Trained entirely by EP, with no backpropagation at any stage, the model reaches 3.64 FID on CIFAR-10, against 3.54 for the same network trained by backprop. To my knowledge, no backprop-free generative model has come close to this quality before. A hardware-friendly version of the architecture, built only from local convolutions and pointwise nonlinearities with no normalisation or attention, reaches and FID of 3.51, so the operations that are hardest to build in analogue hardware turn out not to be needed. Along the way I found why EP’s usable nudging range was so narrow for this objective, and fixed it with a linearised nudge that widens the range by about four orders of magnitude.
The same energy landscapes turn up in statistical physics when describing slow relaxation phenomena in (typically disordered) systems such as:
- Glass (struggling to find its global minimum energy crystal configuration)
- Protein folding (struggling to find their native state)
- Optimisation problems (struggling to find their optimal solution)
Using the Rubik’s Cube as a model disordered system, we found a saddles-to-minima topological crossover in the connectivity of critical points of the energy landscape which controls the existence and onset of glassy dynamics — a structural property of the landscape that decides whether noisy gradient descent gets stuck or finds its way to the global minimum.
What else?
Alongside the PhD, I was a technical project lead in applied machine learning work on risk forecasting at the Alan Turing Institute and UK Ministry of Justice.
I submit my thesis in May 2027, and will be looking for research roles in neuromorphic and/or energy-based machine learning from around then — do get in touch.
News
- PotentialEP is on arXiv: the first backprop-free training of a CIFAR-10 generative model to competitive quality (3.64 FID).
- FlowEqProp wins Best Paper at ICONS ‘26 Award
- Selected for the ELLIS Cambridge Summer School on Probabilistic Machine Learning
- Presented Learning at the Speed of Physics at the NeurIPS ML4PS workshop
- Started as ML Technical Project Lead, Alan Turing Institute & UK Ministry of Justice
Selected publications
- FlowEqProp: Training Flow Matching Generative Models with Gradient Equilibrium PropagationIn Proceedings of the International Conference on Neuromorphic Systems (ICONS ’26), 2026