Associative completion and investment learning using PSOMs

Abstract

We describe a hierarchical scheme for rapid adaptation of context dependent “skills”. The underlying idea is to first invest some learning effort to specialize the learning system to become a rapid learner for a restricted range of contexts. This is achieved by constructing a “Meta-mapping” that replaces an slow and iterative context adaptation by a “one-shot adaptation”, which is a context-dependent skill-reparameterization.

The notion of “skill” is very general and includes a task specific, hand-crafted function mapping with context dependent parameterization, a complex control system, as well as a general learning system.

A representation of a skill that is particularly convenient for the investment learning approach is by a Parameterized Self-Organizing Map (PSOM). Its direct constructability from even small data sets significantly simplifies the investment learning stage; its ability to operate as a continuous associative memory allows to represent skills in the form of “multi-way” mappings (relations) and provides an automatic mechanism for sensor data fusion.

We demonstrate the concept in the context of a (synthetic) vision task that involves the associative completion of a set of feature locations and the task of one-shot adaptation of the transformation between world and object coordinates to a changed camera view of the object.