|
|
About Interests Gallery Resume |
Since the early days of Hubel and
Wiesel, researchers have been trying to understand what individual neurons
signal. Predictive coding has
emerged as a general principle to explain perception, but given the
uniformity of the cortex, it seems likely that the brain uses the same
principle throughout. If this is
the case, then there must exist some principle for which optimal motor
control and optimal perception are special cases. While reinforcement learning could be considered such a
principle, it does not make use of all of the available training signal. Predictive coding, for example, can
train purely off of the input, without regard for an external teacher, but
reinforcement learning cannot.
My research is aimed at finding a principle
that merges predictive coding and some form of optimal decision making. However, there does not yet exist a
general principle in motor control to match predictive coding. Possibilities for this include
curiosity and ability to control the environment, and I am actively exploring
rigorous formulations of these ideas.
|
· Predictive Coding with Temporal Invariance
Jonathan Shaw
TR859, Computer Science Dept., U.
Rochester, March 2005. PDF
· Predictive Action
Jonathan Shaw
TR858, Computer Science Dept., U. Rochester, March 2005. PDF
·
Temporal
Sparse Coding with Spiking Neurons without Synfire Chains
Jonathan Shaw
TR834, Computer Science Dept., U. Rochester, March 2004. PDF