Fortunately for us humans, the brain is a complex network of billions of neurons and trillions of synapses. Unfortunately for us scientists, the brain is a complex network of billions of neurons and trillions of synapses.
In order to understand how complex function emerges from microscopic interactions making up neuronal activity, there are two opposing approaches: top-down and bottom-up modelling. Now, these two terms are often used in different contexts to mean, frankly, anything. Here, I am referring to the opposing sides of computational models, which can be functional (perform useful tasks, like image classification or protein folding prediction) but abstract, and networks which are biologically accurate, but hard to functionalize:
In order to understand how biological networks perform complex tasks, we need to satisfy both ends of the scale: model complex dynamics, while retaining the functionality and explainability found in simpler networks. My current approach is to start off from algorithms on the left side, and bring them closer to realistic models of neurons.
One issue this raises is the weight symmetry or weight transport problem. Algorithms for training of neural networks, such as Error Backpropagation, wake-sleep/contrastive Hebbian learning or certain implementations of learning by Predictive Coding all require synaptic strengths between two areas to be symmetric (same sign, same magnitude). If the brain uses such a learning algorithm, we need to explain how it could obtain such symmetry.
There are plenty of solutions for simple (rate-based) models, but a solution for more realistic models using spikes was needed. Therefore, we have developed Spike-based Alignment Learning (SAL), the spiking sister theory of my previous work PAL. SAL uses the information about asymmetric weights imprinted on the spiking times to recover and stabilize networks. To spare you a lengthy, technical description, I'll just refer you to our preprint, which contains lengthy, technical descriptions of how it works, and why it is pretty nice:
Weight transport through spike timing for robust local gradients
Timo Gierlich, Andreas Baumbach, Akos F. Kungl, Kevin Max, Mihai A. Petrovici
https://arxiv.org/abs/2503.02642