SGER: Learning to Develop Insight: A Sub-symbolic Approach to Learning Transfer

Dan Ventura Investigator)

Brigham Young University
Provo, UT 84602

One of the powerful aspects of inductive learning is the ability for models to generalize
– to perform well on data upon which they were not trained. Unfortunately, this ability has so far been
exhibited in only a very limited fashion in sub-symbolic artificial systems, almost exclusively as generalization
across instances of a single task. One of the limitations of today’s “intelligent” systems is a fragility due to
an inability to generalize across tasks. Learning transfer is the ability for a system to learn one problem
and then to transfer a significant amount of the learned knowledge to a different problem, allowing instant
performance gains and significantly reducing the learning necessary to become proficient on the second
problem. Creativity is required in deciding which prior knowledge to use and how to use it. Symbolic
systems employing some form of analogy for learning transfer have been somewhat successful here; however,
these approaches require a significant amount of specialized domain knowledge and do not generalize. We
propose the use of sub-symbolic approaches to creative problem solving (via learning transfer), trading the
interpretability of symbolic approaches for representational power and generality.