From biological decision making and temporal learning to ultra-edge AI – MicroBrainOnEdge

Centralized AI models are complex, and the computation required to train large AI systems increases tenfold annually. Current foundational models have hundreds of billions of parameters. Trade-offs between computational capability, resources, size, and power consumption (SWaP) will shortly become critical. Current neural architectures with significant fan-out and fan-in and extensive training rely on digital computing architectures that attempt to mimic how nature computes but not how it functions.

Certain animals can make decisions and perform temporal learning using very small neural networks. Insects, for example, can continuously perform challenging inferences, operate, and learn in difficult-to-predict environments with very low energy expenditure and only a few thousand neurons. Understanding the computational principles, architectures, and neuronal circuits of small animals could provide fundamentally new ways to design energy-efficient neural network architectures.

This project belongs to the emerging field of embodied neuroAI. It proposes to exploit recently acquired biological neuronal wiring diagrams and associated behavioral experiments to design new artificial neural networks. We will develop new modeling approaches and numerical simulations to explore the relationship between the biological neural circuits and the bodies of these animals to extract relevant information from the biological neural connectome. We will seek to see how the physical constraints associated with embodied life influence architectural principles of biological neural networks to add these constraints to artificial ones.

We will leverage learned latent representations of biological neural networks and physics-based constraints to provide a new scheme to represent and create highly compressed neural networks and pave the way to new approach to ultra-lean artificial intelligence.