For privacy and security concerns, the need to erase unwanted information from pre-trained vision models is becoming evident nowadays. In real-world scenarios, erasure requests originate at any time from both users and model owners, and these requests usually form a sequence. Therefore, under such a setting, selective information is expected to be continuously removed from a pre-trained model while maintaining the rest. We define this problem as continual forgetting and identify three key challenges. (i) For unwanted knowledge, efficient and effective deleting is crucial. (ii) For remaining knowledge, the impact brought by the forgetting procedure should be minimal. (iii) In real-world scenarios, the training samples may be scarce or partially missing during the process of forgetting. To address them, we first propose Group Sparse LoRA (GS-LoRA). Specifically, towards (i), we introduce Low-Rank Adaptation (LoRA) modules to fine-tune the Feed-Forward Network (FFN) layers in Transformer blocks for each forgetting task independently, and towards (ii), a simple group sparse regularization is adopted, enabling automatic selection of specific LoRA groups and zeroing out the others. To further extend GS-LoRA to more practical scenarios, we incorporate prototype information as additional supervision and introduce a more practical approach, GS-LoRA++. For each forgotten class, we move the logits away from its original prototype. For the remaining classes, we pull the logits closer to their respective prototypes. We conduct extensive experiments on face recognition, object detection and image classification and demonstrate that our method manages to forget specific classes with minimal impact on other classes.

Reasoning stands as a cornerstone of intelligence, enabling the synthesis of existing knowledge to solve complex problems. Despite remarkable progress, existing reasoning benchmarks often fail to rigorously evaluate the nuanced reasoning capabilities required for complex, real-world problemsolving, particularly in multi-disciplinary and multimodal contexts. In this paper, we introduce a graduate-level, multi-disciplinary, EnglishChinese benchmark, dubbed as Reasoning Bench (RBench), for assessing the reasoning capability of both language and multimodal models. RBench spans 1,094 questions across 108 subjects for language model evaluation and 665 questions across 83 subjects for multimodal model testing. These questions are meticulously curated to ensure rigorous difficulty calibration, subject balance, and cross-linguistic alignment, enabling the assessment to be an Olympiad-level multidisciplinary benchmark. We evaluate many models such as o1, GPT-4o, DeepSeek-R1, etc. Experimental results indicate that advanced models perform poorly on complex reasoning, especially multimodal reasoning. Even the top-performing model OpenAI o1 achieves only 53.2% accuracy on our multimodal evaluation. Data and code are made publicly available athttps://evalmodels.github.io/rbench/

We observe a high level of imbalance in the accuracy of different learned classes in the same old task for the first time. This intriguing phenomenon, discovered in replay-based Class Incremental Learning (CIL), highlights the imbalanced forgetting of learned classes, as their accuracy is similar before the occurrence of catastrophic forgetting. This discovery remains previously unidentified due to the reliance on average incremental accuracy as the measurement for CIL, which assumes that the accuracy of classes within the same task is similar. However, this assumption is invalid in the face of catastrophic forgetting. Further empirical studies indicate that this imbalanced forgetting is caused by conflicts in representation between semantically similar old and new classes. These conflicts are rooted in the data imbalance present in replay-based CIL methods. Building on these insights, we propose CLass-Aware Disentanglement (CLAD) as a means to predict the old classes that are more likely to be forgotten and enhance their accuracy. Importantly, CLAD can be seamlessly integrated into existing CIL methods. Extensive experiments demonstrate that CLAD consistently improves current replay-based methods, resulting in performance gains of up to 2.56%.

For privacy and security concerns, the need to erase un-wanted information from pre-trained vision models is becoming evident nowadays. In real-world scenar-ios, erasure requests originate at any time from both users and model owners. These requests usually form a sequence. Therefore, under such a setting, selective information is expected to be continuously removed from a pre-trained model while maintaining the rest. We define this problem as continual forgetting and identify two key challenges. (i) For unwanted knowledge, efficient and effective deleting is crucial. (ii) For remaining knowledge, the impact brought by the forgetting procedure should be minimal. To address them, we propose Group Sparse LoRA (GS-LoRA). Specifically, towards (i), we use LoRA modules to fine-tune the FFN layers in Transformer blocks for each forgetting task independently, and towards (ii), a simple group sparse regularization is adopted, enabling automatic selection of specific LoRA groups and zeroing out the others. GS-LoRA is effective, parameter-efficient, data-efficient, and easy to implement. We conduct extensive experiments on face recognition, object detection and image classification and demonstrate that GS-LoRA manages to forget specific classes with minimal impact on other classes. Codes will be released on https://github.com/bjzhb666/GS-LoRA.

Continual learning (CL) aims to empower models to learn new tasks without forgetting previously acquired knowledge. Most prior works concentrate on the techniques of architectures, replay data, regularization, etc. However, the category name of each class is largely neglected. Existing methods commonly utilize the one-hot labels and randomly initialize the classifier head. We argue that the scarce semantic information conveyed by the one-hot labels hampers the effective knowledge transfer across tasks. In this paper, we revisit the role of the classifier head within the CL paradigm and replace the classifier with semantic knowledge from pretrained language models (PLMs). Specifically, we use PLMs to generate semantic targets for each class, which are frozen and serve as supervision signals during training. Such targets fully consider the semantic correlation between all classes across tasks. Empirical studies show that our approach mitigates forgetting by alleviating representation drifting and facilitating knowledge transfer across tasks. The proposed method is simple to implement and can seamlessly be plugged into existing methods with negligible adjustments. Extensive experiments based on eleven mainstream baselines demonstrate the effectiveness and generalizability of our approach to various protocols. For example, under the class-incremental learning setting on ImageNet-100, our method significantly improves the Top-1 accuracy by 3.2% to 6.1% while reducing the forgetting rate by 2.6% to 13.1%.

Continual learning strives to acquire knowledge across sequential tasks without forgetting previously assimilated knowledge. Current state-of-the-art methodologies utilize dynamic architectural strategies to increase the network capacity for new tasks. However, these approaches often suffer from a rapid growth in the number of parameters. While some methods introduce an additional network compression stage to address this, they tend to construct complex and hyperparameter-sensitive systems. In this work, we introduce a novel solution to this challenge by proposing Memory-Boosted transformer (MoBoo), instead of conventional architecture expansion and compression. Specifically, we design a memory-augmented attention mechanism by establishing a memory bank where the “key” and “value” linear projections are stored. This memory integration prompts the model to leverage previously learned knowledge, thereby enhancing stability during training at a marginal cost. The memory bank is lightweight and can be easily managed with a straightforward queue. Moreover, to increase the model’s plasticity, we design a memory-attentive aggregator, which leverages the cross-attention mechanism to adaptively summarize the image representation from the encoder output that has historical knowledge involved. Extensive experiments on challenging benchmarks demonstrate the effectiveness of our method. For example, on ImageNet-100 under 10 tasks, our method outperforms the current state-of-the-art methods by +3.74% in average accuracy and using fewer parameters.

In computer vision, fine-tuning is the de-facto approach to leverage pre-trained vision models to perform downstream tasks. However, deploying it in practice is quite challenging, due to adopting parameter inefficient global update and heavily relying on high-quality downstream data. Recently, prompt-based learning, which adds the task-relevant prompt to adapt the pre-trained models to downstream tasks, has drastically boosted the performance of many natural language downstream tasks. In this work, we extend this notable transfer ability benefited from prompt into vision models as an alternative to fine-tuning. To this end, we propose parameter-efficient Prompt tuning (Pro-tuning) to adapt diverse frozen pre-trained models to a wide variety of downstream vision tasks. The key to Pro-tuning is prompt-based tuning, i.e., learning task-specific vision prompts for downstream input images with the pre-trained model frozen. By only training a small number of additional parameters, Pro-tuning can generate compact and robust downstream models both for CNN-based and transformer-based network architectures. Comprehensive experiments evidence that the proposed Pro-tuning outperforms fine-tuning on a broad range of vision tasks and scenarios, including image classification (under generic objects, class imbalance, image corruption, natural adversarial examples, and out-of-distribution generalization), and dense prediction tasks such as object detection and semantic segmentation.

The sparse signals provided by external sources have been leveraged as guidance for improving dense disparity estimation. However, previous methods assume depth measurements to be randomly sampled, which restricts performance improvements due to under-sampling in challenging regions and over-sampling in well-estimated areas. In this work, we introduce an Active Disparity Sampling problem that selects suitable sampling patterns to enhance the utility of depth measurements given arbitrary sampling budgets. We achieve this goal by learning an Adjoint Network for a deep stereo model to measure its pixel-wise disparity quality. Specifically, we design a hard-soft prior supervision mechanism to provide hierarchical supervision for learning the quality map. A Bayesian optimized disparity sampling policy is further proposed to sample depth measurements with the guidance of the disparity quality. Extensive experiments on standard datasets with various stereo models demonstrate that our method is suited and effective in different stereo architectures and outperforms existing fixed and adaptive sampling methods under different sampling rates. Remarkably, the proposed method makes substantial improvements when generalized to heterogeneous unseen domains.

Vision Transformer has shown great visual representation power in substantial vision tasks such as recognition and detection, and thus been attracting fast-growing efforts on manually designing more effective architectures. In this paper, we propose to use neural architecture search to automate this process, by searching not only the architecture but also the search space. The central idea is to gradually evolve different search dimensions guided by their E-T Error computed using a weight-sharing supernet. Moreover, we provide design guidelines of general vision transformers with extensive analysis according to the space searching process, which could promote the understanding of vision transformer. Remarkably, the searched models, named S3 (short for Searching the Search Space), from the searched space achieve superior performance to recently proposed models, such as Swin, DeiT and ViT, when evaluated on ImageNet. The effectiveness of S3 is also illustrated on object detection, semantic segmentation and visual question answering, demonstrating its generality to downstream vision and vision-language tasks. Code and models will be available at https://github.com/microsoft/Cream.