A Comparison of Learning Transfer in Networks and Humans

Learning transfer is the improvement in performance on one task having learnt a related task. That the degree of transfer is signi cantly greater in humans than other primates and animals suggests it is a critical component of higher intelligence. One connectionist approach, weight sharing, represents common task knowledge as weighted connections shared by subnetworks dedicated to individual tasks. Although this technique permits transfer, recent analysis has shown that it does not support the same degree of transfer as humans. In this paper, several extensions are outlined, and their theoretical limits compared. The comparison points to a greater role for control mechanisms in connectionist cognitive models.

• Type: Conference
• Year: 1998
• Where: ICONIP
• Authors: Steven Phillips
• Keywords: ICONIP 1998, ICONIP 2007, Represents Common Task, Task Having Learnt, Technique Permits Transfer