AI SCIENCE
It turns out human wiring still wins
Scientists at the University of Surrey have tested a new way to build AI by copying how the human brain connects its neurons.
Their method, described in Neurocomputing, uses Topographical Sparse Mapping, which links each artificial neuron only to nearby or related ones.
This cuts unnecessary connections and keeps accuracy high in systems like generative AI and ChatGPT.
Dr Roman Bauer says this could help lower the huge energy use seen in today’s large models, which can use over one million kilowatt-hours to train.
TL;DR
Fewer, smarter connections make AI more efficient without losing accuracy.
Inspired by how the brain learns and prunes neural links.
Could help future neuromorphic hardware.
Neural feng shui
A newer version, Enhanced Topographical Sparse Mapping, also “prunes” connections during training, similar to how the brain removes weaker links as we learn.
The team is also looking at how this approach could help shape neuromorphic computers, which are designed to work more like real brains.
Human brains are about to be the next H200 chip - MV