EBRAINS User Stories – Julian Göltz: rethinking how spiking neural networks learn

This time, we speak with Julian Göltz, a postdoctoral researcher at the University of Bern. His work focuses on spiking neural networks and neuromorphic computing, exploring new ways to train brain-inspired AI systems more efficiently. In this interview, he talks about his PhD research. 

What question did you set out to answer in your PhD? What challenge were you trying to solve?

From the beginning, I was interested in spiking neural networks – first in themselves, and then in how to make them perform certain tasks from a machine learning perspective. 

When I started my PhD, a method called surrogate gradients had just been established to train spiking neural networks. It works very well but has certain drawbacks: in order to train a Spiking Neural Network (SNN), in a certain way it ‘modifies’ the neurons to make standard ANN-methods applicable – leading to an increase of both compute and memory. 

I wondered whether we could describe spiking neural networks directly in terms of the sparse activity that is actually defining them: the action potentials. Neurons – both biological and simulated – communicate via such discrete spikes at specific points in time. The main result of my thesis was showing that you can describe a neuron's output as a mathematically analytical function of its input, and use this to train SNNs. 

This was the core theoretical contribution. The next question was whether it works in practice, given how finicky spiking networks can be. I tested the method in simulations and then extended it to neuromorphic hardware. 

You introduced a new method for training spiking neural networks. In simple terms, how does it work and what makes it different?

In most neuron models, the central variable is the membrane potential – in a biological neuron this is the voltage difference between the inside and the outside of the cell. Neurons act as leaky integrators: they receive input, creating an excitation that in time decays again, but if enough accumulates, the neurons produce a spike and reset. Spikes are usually represented as delta functions defined only by their timing, and because the membrane potential resets abruptly, there’s a discontinuity that makes gradient-based training difficult. 

The key idea is to focus not on the membrane but on spike timing instead. The timing of a spike depends in a controlled way on the input: if you slightly change the input or the weights, the spike time in turn shifts slightly. We were able to describe this relationship exactly with analytical formulas. Because of this property, you can determine: if you want to change the output spike time, you can calculate how inputs or weights need to change to achieve that. 

Since this operates at the level of individual spikes, the method is inherently event-based. And because the formulas are analytical, the gradients are exact. You know precisely how much to adjust inputs or weights to produce a desired change in output. That’s essentially the novelty. 

You demonstrated this on BrainScaleS-2. What does this system allow you to do, and what makes neuromorphic hardware special? 

There are two main motivations for neuromorphic hardware: understanding the brain by mimicking it, and profiting off its efficiency for applications. The brain uses very little power, and that must be related to how it works. By copying aspects of its dynamics, we hope to gain some of that efficiency. 

BrainScaleS-2 is particularly interesting because it is designed in an analogue fashion: Each neuron on the chip behaves independently, so no matter how many neurons are used, the runtime of a simulation is the same. It’s also accelerated – about 1,000 times faster than biological real time: processes that take seconds in the brain happen in milliseconds on the chip. 

This enables very fast inference. We showed this with a standard dataset, training the network with the chip “in the loop.” Because the hardware is analogue, each circuit behaves slightly differently, so the network has to adapt to these irregularities.  

In a full system test – including communication between the host computer and the chip – we classified all 10,000 images from the test set in under one second. That includes sending the data, running the computation, and receiving the results, all while achieving high accuracy. 

Overall, BrainScaleS-2 leverages time-based computation, like biological neurons, and lets us combine this with our algorithm to achieve fast, energy-efficient inference. 

Looking ahead, how could methods like yours contribute to more efficient AI or brain-inspired technology? 

There are several directions. In our most recent publication, we included transmission delays in the network – which in computational neuroscience are often ignored or appear in a simplified way. In our framework, delays are easy to incorporate: since we work with spike times, a delay is just an addition to the spike time, and the gradient of this operation is straightforward to compute. 

This allowed us to study the computational role of delays in detail and explore how different types affect performance. We also implemented this on BrainScaleS-2 in an ad hoc way and found that it works on hardware and can even stabilise training. 

A natural next step would be designing hardware that includes such delays natively, combining efficient analogue computation with the computational advantages of delays. 

In addition, we are studying how information is processed in our networks: it turns out, that in deeper networks a wave of spikes drives the activity in a certain way that is also seen in the visual cortex. We are analysing this in an ongoing collaboration with the Forschungszentrum Jülich, in an effort to both learn about biological networks and by understanding the information flow in the networks to optimise their capabilities.  

What are your plans for the future? What are you excited about? 

I’ve just moved to Bern – this is my first week in Switzerland – and the weather is so nice outside that it feels a bit unfortunate to be in the office. Going forward, I’m interested in generalising this work. In my thesis, I outlined how the approach could be extended to more general neuron models, activation functions, and network architectures. I am excited to test it in more complex and application-oriented settings, and connect this theory with more bioplausible ideas of how errors are distributed in neuronal networks. 

Read more

EBRAINS User Stories series 

From brains to AI: a new theory of learning through time 

Ultrafast nanolasers mimic how the brain imagines unseen parts of the world 

How the brain encodes uncertainty