Towards Robust and Cost-Efficient Knowledge Unlearning for Large Language Models

Large Language Models (LLMs) have demonstrated strong reasoning and memorization capabilities via pretraining on massive textual corpora. However, training LLMs on human-written text entails significant risk of privacy and copyright violations, which demands an efficient machine unlearning framework to remove knowledge of sensitive data without retraining the model from scratch. While Gradient Ascent (GA) is widely used for unlearning by reducing the likelihood of generating unwanted information, the unboundedness of increasing the cross-entropy loss causes not only unstable optimization, but also catastrophic forgetting of knowledge that needs to be retained. We also discover its joint application under low-rank adaptation results in significantly suboptimal computational cost vs. generative performance trade-offs. In light of this limitation, we propose two novel techniques for robust and cost-efficient unlearning on LLMs. We first design an Inverted Hinge loss that suppresses unwanted tokens by increasing the probability of the next most likely token, thereby retaining fluency and structure in language generation. We also propose to initialize low-rank adapter weights based on Fisher-weighted low-rank approximation, which induces faster unlearning and better knowledge retention by allowing model updates to be focused on parameters that are important in generating textual data we wish to remove.

Towards Realistic Incremental Scenario in Class Incremental Semantic Segmentation

This paper addresses the unrealistic aspect of the commonly adopted Continuous Incremental Semantic Segmentation (CISS) scenario, termed overlapped. We point out that overlapped allows the same image to reappear in future tasks with different pixel labels, which is far from practical incremental learning scenarios. Moreover, we identified that this flawed scenario may lead to biased results for two commonly used techniques in CISS, pseudo-labeling and exemplar memory, resulting in unintended advantages or disadvantages for certain techniques. To mitigate this, a practical scenario called partitioned is proposed, in which the dataset is first divided into distinct subsets representing each class, and then the subsets are assigned to each corresponding task. This efficiently addresses the issue above while meeting the requirement of CISS scenario, such as capturing the background shifts. Furthermore, we identify and address the code implementation issues related to retrieving data from the exemplar memory, which was ignored in previous works. Lastly, we introduce a simple yet competitive memory-based baseline, MiB-AugM, that handles background shifts of current tasks in the exemplar memory. This baseline achieves state-of-the-art results across multiple tasks involving learning numerous new classes.

Salience-Based Adaptive Masking: Revisiting Token Dynamics for Enhanced Pre-training

In this paper, we introduce Saliency-Based Adaptive Masking (SBAM), a novel and cost-effective approach that significantly enhances the pre-training performance of Masked Image Modeling (MIM) approaches by prioritizing token salience. Our method provides robustness against variations in masking ratios, effectively mitigating the performance instability issues common in existing methods. This relaxes the sensitivity of MIM-based pre-training to masking ratios, which in turn allows us to propose an adaptive strategy for `tailored' masking ratios for each data sample, which no existing method can provide. Toward this goal, we propose an Adaptive Masking Ratio (AMR) strategy that dynamically adjusts the proportion of masking for the unique content of each image based on token salience. We show that our method significantly improves over the state-of-the-art in mask-based pre-training on the ImageNet-1K dataset.

Hyperparameters in Continual Learning: a Reality Check

Various algorithms for continual learning (CL) have been designed with the goal of effectively alleviating the trade-off between stability and plasticity during the CL process. To achieve this goal, tuning appropriate hyperparameters for each algorithm is essential. As an evaluation protocol, it has been common practice to train a CL algorithm using diverse hyperparameter values on a CL scenario constructed with a benchmark dataset. Subsequently, the best performance attained with the optimal hyperparameter value serves as the criterion for evaluating the CL algorithm. In this paper, we contend that this evaluation protocol is not only impractical but also incapable of effectively assessing the CL capability of a CL algorithm. Returning to the fundamental principles of model evaluation in machine learning, we propose an evaluation protocol that involves Hyperparameter Tuning and Evaluation phases. Those phases consist of different datasets but share the same CL scenario. In the Hyperparameter Tuning phase, each algorithm is iteratively trained with different hyperparameter values to find the optimal hyperparameter values. Subsequently, in the Evaluation phase, the optimal hyperparameter values is directly applied for training each algorithm, and their performance in the Evaluation phase serves as the criterion for evaluating them. Through experiments on CIFAR-100 and ImageNet-100 based on the proposed protocol in class-incremental learning, we not only observed that the existing evaluation method fail to properly assess the CL capability of each algorithm but also observe that some recently proposed state-of-the-art algorithms, which reported superior performance, actually exhibit inferior performance compared to the previous algorithm.

Sy-CON: Symmetric Contrastive Loss for Continual Self-Supervised Representation Learning

We introduce a novel and general loss function, called Symmetric Contrastive (Sy-CON) loss, for effective continual self-supervised learning (CSSL). We first argue that the conventional loss form of continual learning which consists of single task-specific loss (for plasticity) and a regularizer (for stability) may not be ideal for contrastive loss based CSSL that focus on representation learning. Our reasoning is that, in contrastive learning based methods, the task-specific loss would suffer from decreasing diversity of negative samples and the regularizer may hinder learning new distinctive representations. To that end, we propose Sy-CON that consists of two losses (one for plasticity and the other for stability) with symmetric dependence on current and past models' negative sample embeddings. We argue our model can naturally find good trade-off between the plasticity and stability without any explicit hyperparameter tuning. We validate the effectiveness of our approach through extensive experiments, demonstrating that MoCo-based implementation of Sy-CON loss achieves superior performance compared to other state-of-the-art CSSL methods.

Learning to Unlearn: Instance-wise Unlearning for Pre-trained Classifiers

Since the recent advent of regulations for data protection (e.g., the General Data Protection Regulation), there has been increasing demand in deleting information learned from sensitive data in pre-trained models without retraining from scratch. The inherent vulnerability of neural networks towards adversarial attacks and unfairness also calls for a robust method to remove or correct information in an instance-wise fashion, while retaining the predictive performance across remaining data. To this end, we define instance-wise unlearning, of which the goal is to delete information on a set of instances from a pre-trained model, by either misclassifying each instance away from its original prediction or relabeling the instance to a different label. We also propose two methods that reduce forgetting on the remaining data: 1) utilizing adversarial examples to overcome forgetting at the representation-level and 2) leveraging weight importance metrics to pinpoint network parameters guilty of propagating unwanted information. Both methods only require the pre-trained model and data instances to forget, allowing painless application to real-life settings where the entire training set is unavailable. Through extensive experimentation on various image classification benchmarks, we show that our approach effectively preserves knowledge of remaining data while unlearning given instances in both single-task and continual unlearning scenarios.

Knowledge Unlearning for Mitigating Privacy Risks in Language Models

Pretrained Language Models (LMs) memorize a vast amount of knowledge during initial pretraining, including information that may violate the privacy of personal lives and identities. Previous work addressing privacy issues for language models has mostly focused on data preprocessing and differential privacy methods, both requiring re-training the underlying LM. We propose knowledge unlearning as an alternative method to reduce privacy risks for LMs post hoc. We show that simply applying the unlikelihood training objective to target token sequences is effective at forgetting them with little to no degradation of general language modeling performances; it sometimes even substantially improves the underlying LM with just a few iterations. We also find that sequential unlearning is better than trying to unlearn all the data at once and that unlearning is highly dependent on which kind of data (domain) is forgotten. By showing comparisons with a previous data preprocessing method known to mitigate privacy risks for LMs, we show that unlearning can give a stronger empirical privacy guarantee in scenarios where the data vulnerable to extraction attacks are known a priori while being orders of magnitude more computationally efficient. We release the code and dataset needed to replicate our results at https://github.com/joeljang/knowledge-unlearning .

Is Continual Learning Truly Learning Representations Continually?

Continual learning (CL) aims to learn from sequentially arriving tasks without forgetting previous tasks. Whereas CL algorithms have tried to achieve higher average test accuracy across all the tasks learned so far, learning continuously useful representations is critical for successful generalization and downstream transfer. To measure representational quality, we re-train only the output layers using a small balanced dataset for all the tasks, evaluating the average accuracy without any biased predictions toward the current task. We also test on several downstream tasks, measuring transfer learning accuracy of the learned representations. By testing our new formalism on ImageNet-100 and ImageNet-1000, we find that using more exemplar memory is the only option to make a meaningful difference in learned representations, and most of the regularization- or distillation-based CL algorithms that use the exemplar memory fail to learn continuously useful representations in class-incremental learning. Surprisingly, unsupervised (or self-supervised) CL with sufficient memory size can achieve comparable performance to the supervised counterparts. Considering non-trivial labeling costs, we claim that finding more efficient unsupervised CL algorithms that minimally use exemplary memory would be the next promising direction for CL research.

Task-Balanced Batch Normalization for Exemplar-based Class-Incremental Learning

Batch Normalization (BN) is an essential layer for training neural network models in various computer vision tasks. It has been widely used in continual learning scenarios with little discussion, but we find that BN should be carefully applied, particularly for the exemplar memory based class incremental learning (CIL). We first analyze that the empirical mean and variance obtained for normalization in a BN layer become highly biased toward the current task. To tackle its significant problems in training and test phases, we propose Task-Balanced Batch Normalization (TBBN). Given each mini-batch imbalanced between the current and previous tasks, TBBN first reshapes and repeats the batch, calculating near task-balanced mean and variance. Second, we show that when the affine transformation parameters of BN are learned from a reshaped feature map, they become less-biased toward the current task. Based on our extensive CIL experiments with CIFAR-100 and ImageNet-100 datasets, we demonstrate that our TBBN is easily applicable to most of existing exemplar-based CIL algorithms, improving their performance by decreasing the forgetting on the previous tasks.

Supervised Neural Discrete Universal Denoiser for Adaptive Denoising

We improve the recently developed Neural DUDE, a neural network-based adaptive discrete denoiser, by combining it with the supervised learning framework. Namely, we make the supervised pre-training of Neural DUDE compatible with the adaptive fine-tuning of the parameters based on the given noisy data subject to denoising. As a result, we achieve a significant denoising performance boost compared to the vanilla Neural DUDE, which only carries out the adaptive fine-tuning step with randomly initialized parameters. Moreover, we show the adaptive fine-tuning makes the algorithm robust such that a noise-mismatched or blindly trained supervised model can still achieve the performance of that of the matched model. Furthermore, we make a few algorithmic advancements to make Neural DUDE more scalable and deal with multi-dimensional data or data with larger alphabet size. We systematically show our improvements on two very diverse datasets, binary images and DNA sequences.