Summary
A research core: spiking networks, local learning, grown rather than hand-designed structure. The source is out — architecture is encoded as a compact genome and grown by a developmental process, and you can read exactly how.
MagicBrain is the laboratory’s own line of research into brain-like computation: impulse-based networks, learning rules that stay local to the connection, and network structure that develops instead of being designed by hand.
It is published rather than described: the library is on GitHub under AGPL-3.0 with a commercial alternative, and a companion playground runs it for real — training with held-out evaluation, neurogenesis stage by stage, damage and repair, and a comparison against the compensated arithmetic of Balansis. The numbers in that playground are generated from actual runs, including the ones that do not flatter the approach.
Links