Generative Brain Model virtual whole zebrafish brain with neuron spike fidelity

Thomson Laboratory Caltech
github.com/Sam-Fatehmanesh/GBM
Background Background

Architecture

Neural Spike Conversion: Clean representation eliminating noise and artifacts

Mamba Architecture: Separates observation from hidden state dynamics for interpretability

Two-Phase Training: Pretrain across many fish, then fine-tune on individual brains

Left Image Credit
Background

Future Steps

Closed-Loop Enhancement: Precise prediction and replication of cognitive functions

Fast Think System: Predict neural activity faster than the zebrafish generates it, enabling super reaction speeds

Brain LoRA: Transfer novel abilities and sensory functions from model to living brain