Failures

Catastrophic Forgetting

1 min read

Quick Definition

When an agent loses previously learned capabilities after being trained on new tasks or data.

Neural networks can overwrite old knowledge when learning new things, a challenge for continual learning.

Manifestations

  • Lost capabilities after fine-tuning
  • Degraded general knowledge
  • Skill regression

Mitigation

  • Elastic weight consolidation
  • Experience replay
  • Multi-task training
  • Careful fine-tuning scope
failurestraininglearning