Graph neural networks uncover structure and functions underlying the activity of simulated neural assemblies
This new method could finally crack the black box of neural activity...
Deep Dive
Researchers have developed a graph neural network (GNN) framework that can reverse-engineer the inner workings of simulated neural assemblies containing thousands of neurons. Unlike opaque models like RNNs and transformers, this method not only predicts neural activity but also interprets it, revealing the underlying connectivity matrix, neuron types, signaling functions, and even hidden external stimuli. The breakthrough offers a new tool for decomposing complex, heterogeneous systems into simple, interpretable representations.
Why It Matters
It provides a powerful, interpretable tool for neuroscience that could accelerate our understanding of real brain circuits and complex systems.