Context-sensitive multiple task learning with consolidated domain knowledge
LE3 .A278 2011
2011
Silver, Danny
Acadia University
Master of Science
Masters
Computer Science
Machine lifelong learning (ML3) is concerned with machines capable of learning and retaining knowledge over time, and exploiting previous knowledge to assist new learn- ing. An ML3 system requires e ective task retention, and e ective consolidation of new tasks. This thesis presents an ML3 system using a context-sensitive multiple task learning (csMTL) neural network that functions as a consolidated domain knowledge store. csMTL was developed in response to structural limitations of multiple task learning (MTL) for ML3. Instead of additional outputs for each task csMTL uses ad- ditional context inputs that indicate the associated task. The csMTL-based system is analyzed empirically using synthetic and real domains. The experiments focus on the e ective retention of knowledge and the e ective consolidation of new knowledge, us- ing independent test set accuracy as a measure of e ectiveness. The studies indicate that the methodology results in e ective task retention when appropriate learning parameters are used. New task consolidation e cacy su ered using the same learn- ing parameters. Experimentation also suggests that virtual instances (input-output pairs constructed from the consolidated domain knowledge) that correspond to real instances improves e cacy of retention and new task consolidation. Experimentation also indicates that representational transfer allows more e ective retention, but at the cost of less e ective new task consolidation.
The author retains copyright in this thesis. Any substantial copying or any other actions that exceed fair dealing or other exceptions in the Copyright Act require the permission of the author.
https://scholar.acadiau.ca/islandora/object/theses:175