Speaker: Jascha Achterberg ARNI WG Multi-resource-cost optimization of neural network models

Title:
Building the brain’s efficient system-level architecture: optimisations across space, time, and multiple regions

Abstract:
The computations a brain can perform are fundamentally constrained by physical realities: energetic resources are limited, and time is precious. To understand why the brain works the way it does, we must understand its function in the context of these constraints. Prior modeling work has successfully demonstrated how spatial energetic constraints drive structure-function co-optimization, giving rise to many of the architectural features we observe across areas of neuroscience. By incorporating such physical constraints, we can build complex systems-level models that are meaningfully constrained by physically measurable factors rather than arbitrary design choices.

In this talk, I will expand on these spatial frameworks by introducing new work on temporal processing and signal precision constraints in neural networks. I will demonstrate how different optimization strategies within individual regions can be combined in heterogeneous multi-region models, revealing how the brain trades off resource use across tasks and situations. Finally, I will show how space and time interact in surprising ways to achieve efficient computation — principles that apply not only to the brain but to any large-scale distributed computing system. Together, these advances bring us closer to understanding the general principles that enable sophisticated intelligence to emerge from physically and energetically constrained computing systems.