Shultz, T.
Department of Psychology McGill University, 1205 Penfield Ave., Montreal, Quebec, Canada H3A 1B1
Autonomous learning is the ability to learn effectively without much external assistance. As such, autonomy is a desired quality in fields such as machine learning and artificial intelligence where the effectiveness of learning systems is seriously compromised whenever human intervention is required. It is likewise a critical feature in cognitive science where the idea is to understand the adaptive learning of human and other biological agents in their natural habitats. An important strength of autonomous learners is that they can shape their own learning and development, in large part by choosing which problems to learn. Such choices include selecting a problem to learn and deciding whether to continue learning on that selected task or abandon it in favor of something else.
Here, we extend a constructive neural-learning algorithm, sibling-descendant cascade-correlation, to monitor lack of progress in learning so that unproductive learning can be abandoned. The algorithm extension defines a learning cycle as a combination of an input phase (which recruits a new hidden unit to afford reconceptualization of the problem) and an adjacent output phase (which incorporates the new hidden unit into efforts to reduce error). A new, outside loop monitors progress across learning cycles with the aid of two parameters (threshold and patience), just as in the inside input- and output-phase loops. Learning is abandoned when network error fails to change by more than the threshold for patience consecutive learning cycles. The extended algorithm simulates results of recent experiments with infants who abandon learning on difficult tasks and focus their attention on tasks of moderate difficulty (the so-called Goldilocks effect). Moreover, it also avoids network overtraining effects in a more realistic and autonomous manner than conventional use of unrealistic validation test sets. The contributions and current limitations of constructive neural networks for achieving autonomy in learning are briefly assessed.