Continual Learning Metrics · GEM

Gradient Episodic Memory for Continual Learning

David Lopez-Paz, Marc'Aurelio Ranzato · NeurIPS 2017

Read paper · arXiv:1706.08840

Compute accuracy and forward/backward transfer from a task evaluation matrix, separating transfer from peak-based forgetting.

What you will implement

Implement GEM's ACC, BWT and FWT plus an explicitly defined peak-forgetting extension; GEM gradient projection is outside this exercise.

• 96. Continual Learning: Evaluation Matrices and Forgetting

  Compute accuracy, forward/backward transfer and peak forgetting from a stage-by-task evaluation matrix.

  Implement GEM's three evaluation metrics plus an explicit peak-forgetting extension; not GEM gradient projection or a complete continual-learning algorithm.