Do Not Forget to Attend to Uncertainty while Mitigating Catastrophic Forgetting
This work aims to mitigate catastrophic forgetting for deep learning models in incremental learning scenarios, offering an incremental improvement over existing knowledge distillation methods.
This paper addresses catastrophic forgetting in incremental learning by incorporating uncertainty estimation from older task models, which previous knowledge distillation methods often overlook. By using a Bayesian formulation for data and model uncertainties and a self-attention framework, the authors define distillation losses based on aleatoric uncertainty and self-attention, leading to improved accuracy on standard benchmarks.
One of the major limitations of deep learning models is that they face catastrophic forgetting in an incremental learning scenario. There have been several approaches proposed to tackle the problem of incremental learning. Most of these methods are based on knowledge distillation and do not adequately utilize the information provided by older task models, such as uncertainty estimation in predictions. The predictive uncertainty provides the distributional information can be applied to mitigate catastrophic forgetting in a deep learning framework. In the proposed work, we consider a Bayesian formulation to obtain the data and model uncertainties. We also incorporate self-attention framework to address the incremental learning problem. We define distillation losses in terms of aleatoric uncertainty and self-attention. In the proposed work, we investigate different ablation analyses on these losses. Furthermore, we are able to obtain better results in terms of accuracy on standard benchmarks.