Topology Learning

Cellular systems often need to know the physical adjacency of their components to process real-world inputs in a sensible way. This adjacency can be specified beforehand, which is often laborious and not plausible for implementation in biological systems, or it can be learned. We have developed an algorithm for learning the topology of event-based systems which uses the timing of incoming events to determine adjacency, using a modified form of spike-timing dependent plasticity (STDP).

Below is a short video of the algorithm learning the adjacency of interactive floor tiles used in the Ada project from about 90 minutes of real pedestrian input. We have also implemented the algorithm on a silicon retina to learn pixel adjacency from real visual input.

Click here for the paper (Boerlin et al., Neural Computation 21(1):216-238, 2009).