Is spike timing essential to cortical computation, or just an efficient way to transmit information? In this interview I discuss moving from modelling cortical computation in academia to engineering spiking neural networks at SynSense; the search for general principles in neocortex; and what a complete connectome would and would not tell us.

Dylan Muir is a computational neuroscientist (PhD, ETH Zürich), former VP for Global Research Operations at SynSense, and founder and CEO of LexChip.

You moved from modelling cortical computation in academia to building spiking neural networks at SynSense. Did you see SNNs as a model of how brains compute, or as an engineering tool?

It's difficult to say whether the spiking aspect of biological neurons is an essential part of their computation, or if it's really a way to efficiently transmit information along a cell membrane, and therefore between cells. There's certainly some evidence that precise spike timing carries information; there's a very nice paper from Sejnowski showing very precise, very reliable spike timing in some regimes (Mainen & Sejnowski, 1995) [1]. But there is also clearly evidence of information encoded by spike frequency from individual neurons. Most of the work done on receptive fields in visual cortex works off either firing rates of cells, or firing rates of large populations of neurons.

During my time in academia, I built models of cortical computation that were both firing rate models and spike timing models. I saw the patterns of connectivity and the rules for connectivity as more important, compared with whether or not neurons were spiking or firing-rate neurons. On top of that, we know that there are neurotransmitters such as dopamine where it's not completely clear whether precise timing is signal or not.

When I moved to SynSense, the goal was to engineer efficient solutions to particular problems, without any reference to biology per se. It didn't matter whether or not our network architectures resembled cortex, as long as they were a functional engineering solution to whatever application we were trying to build. The point of neocortex, and trying to understand neocortex, was to find simple patterns, or at least comprehensible patterns and logic, that gave rise to information processing in biological nervous systems, and which could lead to a general computational principle used by neocortex to solve a large range of sensory and motor problems.

I can't say whether or not individual spike timing is crucial computationally for cortex. I suspect that evolutionary principles mean that any niche approach for information encoding can and will be used by neurons, if it gives rise to some efficiency.

If neurons will use any encoding trick that works, does that undercut the search for general principles in neocortex?

The argument for a general principle in neocortex comes from the regular architecture, in terms of types of neurons, layers of neurons, and general connectivity rules that seem to exist in vastly different parts of neocortex. For subcortical modules, for example brainstem modules, and for other animals such as birds and insects which do not have neocortex, there could be a much larger degree of hardwiring, and a much looser adherence to any particular coding principle.

Since neocortex has expanded very rapidly in evolutionary time in higher primates, and since the structure looks so repeated and reused from visual cortex to prefrontal cortex, the hope is that there is indeed a general principle or principles which can be learned. So my hope would be that coding principles are likewise conserved across different areas of neocortex, in a way that may not hold for other, older, hardwired modules.

Did any of your academic work on connectivity make its way into SynSense designs?

In the early days at SynSense there was indeed some overlap, since I had done work on efficient encoding and computation in semi-random networks. As we progressed at SynSense, we found that relying on random network connectivity was fairly inefficient, in terms of numbers of neurons and synapses needed, and probably not optimal computationally. We were always focused on very small networks of neurons, meaning that hardwiring is an achievable approach, and maybe even a more optimal approach.

The utility of a repeated general principle is that it's cheap to encode genetically. That constraint isn't one which applies so readily to the large neural network architectures of 2026, since we are very happy to use enormous quantities of storage to specify their weights. Those large amounts of memory are also needed during training, as the learning rules we are using require complete knowledge of the entire network simultaneously. I'm referring to backpropagation, of course.

During development of neocortex, and brains in general, you do not have anywhere near enough information in the genome to encode every neuron, every synapse, every connection. That means it's a much more efficient approach to come up with a general principle and then just replicate that out a large number of times. You then rely on fine-tuning, biological learning, and other genetic principles to control migration and global structure of the brain, rather than backpropagation as a learning rule requiring global knowledge over the entire cortex.

Some researchers argue that gradient descent occurs in biological neural tissue. Do you agree?

There are some researchers, for example Konrad Kording and colleagues, who do believe that gradient descent occurs in biological neural tissue during learning (Richards et al., 2019) [2]. I don't find those arguments very convincing, simply because any learning rule where a performance metric can be defined must minimise that performance metric. Therefore, any learning rule with a measurable goal can be said to be “gradient descent” in some respect. To me that is not the same as synapses and neurons explicitly computing the gradient of a loss function with respect to their own neural activity or synaptic weight.

Where did the engineering at SynSense end up?

Towards the end at SynSense, we found it very efficient to use small network architectures, use backpropagation-style gradient descent algorithms running on GPUs, and end up with hardwired network architectures to perform a single task. Since we were not working in a computational neuroscience research domain, we didn't spend any time analysing the structure of those networks to understand the trained architectures that resulted. We were not trying to speak from trained spiking neural networks back to biological nervous tissue; nor were we trying to apply aspects of biological neural architecture to engineered spiking neural network applications.

If you had a complete wiring diagram of a brain, would you understand it?

The thought experiment in neuroscience was always: if you had access to the entire connectome of a brain, would you automatically understand the computational principles? The answer, in my opinion, is that once you have the full connectome, then your reverse engineering work can begin. The benefit would be an accurate wiring diagram to explore, as opposed to the experimental situation when I was working as an active neuroscience researcher, where you have very sparse and incomplete information about connectivity between neurons in a brain.

Improvements in electron microscopy, and very large scale reconstruction of electron microscopy sections, have meant that we can have connectivity diagrams to some degree of accuracy for, say, a cubic millimetre of cortex (MICrONS Consortium, 2025) [3]. But again, that doesn't automatically give you insight. From this point we can begin to look for connectivity rules, and understand the computational principles implied by those connectivity rules. That's exactly the approach I was taking during my time in academia: trying to define connectivity rules based on function, and define the computational principles that arose from those connectivity rules.

References

[1] Mainen, Z. F., & Sejnowski, T. J. (1995). Reliability of spike timing in neocortical neurons. Science, 268(5216), 1503–1506. https://doi.org/10.1126/science.7770778

[2] Richards, B. A., et al. (2019). A deep learning framework for neuroscience. Nature Neuroscience, 22, 1761–1770. https://doi.org/10.1038/s41593-019-0520-2

[3] The MICrONS Consortium (2025). Functional connectomics spanning multiple areas of mouse visual cortex. Nature, 640(8058), 435–447. https://doi.org/10.1038/s41586-025-08790-w