Hypothesis
Network structure that develops — the way biological tissue does — can reach behaviours that hand-designed architectures reach only by brute force.
Problem
Architecture search treats structure as a hyperparameter. Biology treats it as the product of development. Those are very different search spaces.
Current state
Maturing — mechanism published.
Methods
- Spiking (impulse-based) network substrate
- Learning rules local to the connection
- Structural development instead of fixed topology
Findings
- Local learning plus developing structure trains at research scale, and the mechanism is now open: the library is on GitHub under AGPL-3.0
- Measured honestly, the substrate beats a unigram baseline on character-level text but does not reach a bigram model — the runs and the numbers are published alongside the playground