Year: 2005
Type: article
Abstract: In training two networks on tasks that are different on the surface, but similar, or even isomorphic, at a higher level of description, similarities between the network solutions are plausible if not expected.Such similarities tend to become evident when two networks with shared weights are trained on similar tasks.After training, the shared weights were used as part of a third network that was trained on a third task similar to the first two.This "head start" results in significantly shorter training times than a network that starts with random weights.Shared hidden unit response profiles were analyzed across networks trained on structurally analogous tasks to reveal parallel, but nonidentical features.
Authors: Jianghua Bao, Paul Munro
Institution: University of Pittsburgh
Language: English
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
Funder: University of Pittsburgh
OpenAlex ID: W2590487015