Consolidation using context-sensitive multiple task learning

摘要: Machine lifelong learning (ML3) is concerned with machines capable of and retaining knowledge over time, exploiting this to assist new learning. An ML3 system must accurately retain prior tasks while consolidating in tasks, overcoming the stability-plasticity problem. A presented using a context-sensitive multiple task (csMTL) neural network. csMTL uses single output additional context inputs for associating examples tasks. csMTL-based analyzed empirically synthetic real domains. The experiments focus on effective retention consolidation both functional representational transfer. results indicate that combining two methods transfer serves best knowledge, but at cost less consolidation.