Continual Unlearning (CUL)

Continual Unlearning (CUL) is a machine learning paradigm which allows continual learning to unlearn previous trained tasks at the end of any task’s training. Please refer to my slides about continual unlearning.

Definition

In CLArena, continual unlearning is specifically designed for classification problems and follows the formal definition below. Note that the continual learning part of this paradigm is the same as Continual Learning (CL).

Definition 1 (Continual Unlearning Classification) Given:

• An initialized neural network model $f$ consisting of:
  • A shared backbone network $B$
  • Task-specific output heads
• Sequential tasks: $t=t_1,t_2,\dots$ with task IDs $t_k\in {\mathbb{N}}^{+}$
• For each task $t$, we have:
  • Training data: ${\mathcal{D}}_{\text{train}}^{(t)}=\{({\mathbf{x}}_i,y_i){\}}_{i=1}^{N_t}\in ({\mathcal{X}}^{(t)},{\mathcal{Y}}^{(t)})$
  • Validation data: ${\mathcal{D}}_{\text{val}}^{(t)}\in ({\mathcal{X}}^{(t)},{\mathcal{Y}}^{(t)})$
  • Test data: ${\mathcal{D}}_{\text{test}}^{(t)}\in ({\mathcal{X}}^{(t)},{\mathcal{Y}}^{(t)})$
  • Unlearning request: $u(t)$, a set of task IDs to unlearn after training task $t$

Objective: Develop an algorithm that updates the model from $f^{(t-1)}$ to $f^{(t)}$ when learning task $t$, and an unlearning algorithm that unlearns tasks in $u(t)$, such that:

• Only current task data ${\mathcal{D}}_{\text{train}}^{(t)}$ and ${\mathcal{D}}_{\text{val}}^{(t)}$ are accessible
• Effectively remove the knowledge of specific seen tasks $U(t)={\cup }_{\tau =1}^{t}u(\tau )$ from the model after training task $t$
• Maintain good performance on test datasets of all remaining tasks: $R(t)=\{t_1,\cdots ,t\}-U(t)$

Supported Pipelines

CLArena supports the following experiment and evaluation pipelines for continual learning:

• Continual Unlearning Main Experiment: The primary experiment for training and evaluating continual unlearning models. See Continual Unlearning Main Experiment.
  • Continual Unlearning Main Evaluation: The evaluation phase for assessing the performance of the trained continual unlearning models. Since unlearning evaluation cannot be done with main model only, continual unlearning main evaluation is the same as continual learning. See Save and Evaluate Model (CL Main).
• Continual Unlearning Full Experiment: A comprehensive experiment that evaluates more metrics (including unlearning performance) for continual unlearning. This involves the main experiment, additional reference experiments and continual unlearning full evaluation based on these results. See Continual Unlearning Full Experiment.
  • Reference Retrain Experiment (Continual Unlearning): The reference retrain experiment for full evaluation. See Reference Retrain Experiment.
  • Reference Original Experiment (Continual Unlearning): The reference original experiment for full evaluation. See Reference Original Experiment.
  • Continual Unlearning Full Evaluation: The evaluation phase of the full experiment. See Full Evaluation.