Incremental Slow Feature Analysis

Incremental Slow Feature Analysis (IncSFA) is a biologically plausible, covariance-free method to update slow features, having linear complexity with respect to input dimensionality (batch SFA’s updating complexity is cubic).

IncSFA is amenable to processing of high-dimensional visual input streams, useful for reinforcement learning, and appropriate for developmental robots. Slow feature representations enable a tractable state-space for reinforcement learning on high-dimensional observations. Moreover, the incrementally-computed slow features have formally been shown to be generally useful for RL, as they are approximations of proto-value functions. For example, see the IncSFA-TD technique, which combines IncSFA and temporal difference learning. And since we can compute the features incrementally, we can measure the features' learning progress - applicable to artificial curiosity, an important aspect of developmental robotics. See on-going work on Curious Dr. MISFA.

Hierarchical SFA, combined with competitive learning, has been shown to learn place cells, head-direction cells, and grid cells, from high-dimensional visual input. We developed a Hierarchical IncSFA. Results show that the top layer IncSFA units indeed had learned to code for position, and the responses of several competitively learned units operate as place cells.


Contact ---,