Wrapping up October with a milestone that was both a first and a privilege: supervising my first master's dissertation for Joe McGlynn, who I’m very proud to say has just received a Distinction grade for his excellent work.
The project I proposed focused on developing new training paradigms for spiking neural networks (SNNs)—a new, fascinating and rapidly evolving area of AI that draws inspiration directly from the brain, modelling discrete neuron spikes instead of continuous activation. This shift introduces significant complexity, as traditional differentiability required for model training no longer applies. While this presents a formidable challenge for researchers, the potential rewards are substantial, with SNNs offering faster inference speeds and dramatically lower energy consumption—making it a puzzle well worth solving.
?
If you're curious about why SNNs are so promising and fascinating along with the challenges they present, keep an eye on the Bayezian blog where we’ll soon be publishing an article delving deeper into the potential of Spiking Neural Networks. Myself and the Bayezian team would also love to hear from you on the topic, too.?
?
With my roots in computational neuroscience and current work in computer vision, this project was a perfect intersection of both fields. Working closely with Joe was incredibly rewarding, who approached this challenging topic with incredible dedication and intelligence, making every Friday meeting a weekly highlight.
His novel approach—introducing an automated smoothing relaxed Bernoulli function for stochastic neurons—was equally promising as it was unexplored. This project opened up an incredibly interesting investigation of model evaluation and inference capabilities for this new SNN model.
?
Through our investigations using two event-based vision datasets (N-MNIST and DVS hand gestures), Joe uncovered insights into the robustness of stochastic SNNs that many in the field have overlooked. It was inspiring to see his intellectual creativity and determination from start to finish.
?
Being at the forefront of a rapidly developing and complex field, one I am passionately interested in, working within my domain of computer vision, and to do so with someone as rewarding as Joe to work with has been an absolute pleasure. One I hope to partake in again and I will definitely keep my eyes tightly glued to the world of spiking neural networks. I suggest you do as well (if you can bear the maths).
As mentioned, look out for future articles from Bayezian talking about the promise and challenges of spiking neural networks.
