Visual Analytics on Network Forgetting for Task‐Incremental Learning

Task-incremental learning (Task-IL) aims to enable an intelligent agent continuously accumulate knowledge from new tasks without catastrophically forgetting what it has learned in the past. It drawn increasing attention recent years, with many algorithms being proposed mitigate neural network forgetting. However, none of existing strategies is able completely eliminate issues. Moreover, explaining and fully understanding how forgotten during incremental process still remains under-explored. In this paper, we propose KnowledgeDrift, a visual analytics framework, interpret three objectives: (1) identify when fails memorize past knowledge, (2) visualize information been forgotten, (3) diagnose attained model interferes one Our analytical framework first identifies occurrence by tracking task performance under then provides in-depth inspections drifted via various levels data granularity. KnowledgeDrift allows analysts developers enhance their compare different algorithms. Three case studies are conducted paper further provide insights guidance for users effectively catastrophic over time.