Curriculum-Based Learning with STiNT
Whereas much work in machine learning addresses how to learn a single
function in isolation, I have been interested in how to boost learning
accuracy when multiple functions are learned over time. In this talk
I will address two coupled questions: (1) what algorithms can be
developed that benefit from previous learning when presented with a
new function to be learned? and (2) what curriculum, or ordering of
functions, produces the most effective learning? (This notion of
curricula occurs naturally in situations such as robot learning).
I'll outline a space of algorithms that can benefit from curricula,
and will examine one in detail: STiNT. STiNT combines multiple neural
networks in a directed acyclic graph, so as to benefit from previously
learned functions. STiNT extends that graph as new functions to be
learned arrive so as to approximately maximize the benefit from the
curriculum. By exploring the performance of random, optimal, maximally
pessimistic, and hand-coded curricula, I'll show that both that using
a curriculum is a good thing to do, although the accuracy of learning
with STiNT is strongly influenced by the curriculum used. A neat thing
about STiNT is that a simple greedy algorithm can efficiently use
explicit domain theories and statistical tests to automatically
generate a curriculum that approximates the performance of the optimal
curriculum.
* (This is joint work with Tom Mitchell).