The group had published A Theory of Natural Intelligence, proposing a mechanism for how intelligence emerges in biological learners: self-organising net fragments that assemble into larger structures. The goal of this fellowship was to turn that concept into a technical implementation inside contemporary deep neural networks.
My contribution became the Cooperative Network Architecture, which replaces the distributed activation patterns deep networks rely on with structured, recurrently connected assemblies built from overlapping fragments learned without supervision. Because those fragments compose by construction, the architecture completes occluded figures, tolerates noise and handles out-of-distribution patterns without retraining.
It was published as the cover article of Neural Computation 38(4), pages 538–572.