Class-Incremental Continual Learning into the eXtended DER-verse

Abstract

The staple of human intelligence is the capability of acquiring knowledge in a continuous fashion. In stark contrast, Deep Networks forget catastrophically and, for this reason, the sub-field of Class-Incremental Continual Learning fosters methods that learn a sequence of tasks incrementally, blending sequentially-gained knowledge into a comprehensive prediction. This work aims at assessing and overcoming the pitfalls of our previous proposal Dark Experience Replay (DER), a simple and effective approach that combines rehearsal and Knowledge Distillation. Inspired by the way our minds constantly rewrite past recollections and set expectations for the future, we endow our model with the abilities to i) revise its replay memory to welcome novel information regarding past data ii) pave the way for learning yet unseen classes. We show that the application of these strategies leads to remarkable improvements; indeed, the resulting method -termed eXtended-DER (X-DER) -outperforms the state of the art on both standard benchmarks (such as CIFAR-100 and miniImageNet) and a novel one here introduced. To gain a better understanding, we further provide extensive ablation studies that corroborate and extend the findings of our previous research (e.g. the value of Knowledge Distillation and flatter minima in continual learning setups). We make our results fully reproducible; the codebase is available at https://github.com/aimagelab/mammoth.

R-STAR: Robust Self-Taught Task-Wise Reweighting for Rehearsal-Based Class Incremental Learning

“…Among all approaches for continual learning, the rehearsal-based approaches are the most widely used due to their convenience and effectiveness. In order to tackle the catastrophic forgetting issue, rehearsal-based approaches [27,11,9,2,19,31,22,24,26,7,1,5,8,6,10,4] attempt to preserve the old knowledge from all previous tasks by a memory buffer and replay it when the model learns on a new task. [27] propose to save a small subset of old data and replay it in new tasks.…”

“…Concretely, rehearsal-based approaches attempt to preserve the old knowledge by a continuously updated memory buffer where limited samples are selected from all previous tasks and then available in the new task. Among recent advances of rehearsal-based methods, most of them focus on designing a proper updating strategy of the buffer [26,27,19,6] and exploiting the buffer data with extra regularization constraints [26,7,6,24,6]. Although these efforts indeed contribute to mitigating catastrophic forgetting, the main cause of this challenge for rehearsal-based CIL is still not fully addressed.…”

“…As some recent works [31,1,6] point out, the main cause of catastrophic forgetting for rehearsal-based CIL is the classification score bias. Specifically, the bias is due to data imbalance between old data in the memory buffer and new data in the current task.…”

“…When the model learns on such unbalanced data, it tends to give higher classification scores for new classes. Regarding this, recent methods [31,1,6] propose new strategies to solve the scores bias. [31] train a linear model on a balanced dataset after each task and use it to correct the classification score bias in the last fully connected layer.…”

Effective and efficient continual learning

“…DualHSIC is a general framework that can be combined with almost any mainstream rehearsal-based methods. Therefore, we incorporate DualHSIC into multiple SOTA rehearsal-based methods, including ER [26], DER++ [19], X-DER-RPC [16], and ER-ACE [20],…”