Knowledge Accumulation in Continually Learned Representations and the Issue of Feature Forgetting

Continual learning research has shown that neural networks suffer from catastrophic forgetting "at the output level", but it is debated whether this is also the case at the level of learned representations. Multiple recent studies ascribe representations a certain level of innate robustness against forgetting - that they only forget minimally and no critical information. We revisit and expand upon the experiments that revealed this difference in forgetting and illustrate the coexistence of two phenomena that affect the quality of continually learned representations: knowledge accumulation and feature forgetting. Carefully taking both aspects into account, we show that, even though it is true that feature forgetting can be small in absolute terms, newly learned information tends to be forgotten just as catastrophically at the level of the representation as it is at the output level. Next we show that this feature forgetting is problematic as it substantially slows down knowledge accumulation. Finally, we study how feature forgetting and knowledge accumulation are affected by different types of continual learning methods.