Abrupt representation change refers to the sudden and disruptive modification of a machine learning model internal feature representations when it encounters new, previously unseen data distributions or classes in a sequential learning stream. In continual learning environments, as a neural network attempts to rapidly accommodate unfamiliar concepts alongside historical data, the newly introduced features can heavily interfere or overlap with existing representations, triggering severe parameter adjustments. This rapid reorganization of the latent feature space destabilizes the stability of previously acquired representations, significantly contributing to catastrophic forgetting and impairing the model ability to retain past knowledge while integrating new tasks.
