Lama Alssum

Lama Alssum, Hani Itani, Hasan Abed Al Kader Hammoud et al.

The safety alignment of large language models (LLMs) is becoming increasingly important with their democratization. In this paper, we study the safety degradation that comes with adapting LLMs to new tasks. We attribute this safety compromise to catastrophic forgetting and frame the problem of preserving safety when fine-tuning as a continual learning (CL) problem. We consider the fine-tuning-as-a-service setup where the user uploads their data to a service provider to get a customized model that excels on the user's selected task. We adapt several CL approaches from the literature and systematically evaluate their ability to mitigate safety degradation. These include regularization-based, memory-based, and model merging approaches. We consider two scenarios, (1) benign user data and (2) poisoned user data. Our results demonstrate that CL approaches consistently achieve lower attack success rates than standard fine-tuning. Among these, DER outperforms both other CL methods and existing safety-preserving baselines while maintaining task utility. These findings generalize across three downstream tasks (GSM8K, SST2, Code) and three model families (LLaMA2-7B, Mistral-7B, Gemma-2B), establishing CL as a practical solution to preserve safety.

Lama Alssum, Hasan Abed Al Kader Hammoud, Motasem Alfarra et al.

Deep neural networks suffer from catastrophic forgetting, where performance on previous tasks degrades after training on a new task. This issue arises due to the model's tendency to overwrite previously acquired knowledge with new information. We present a novel approach to address this challenge, focusing on the intersection of memory-based methods and regularization approaches. We formulate a regularization strategy, termed Information Maximization (IM) regularizer, for memory-based continual learning methods, which is based exclusively on the expected label distribution, thus making it class-agnostic. As a consequence, IM regularizer can be directly integrated into various rehearsal-based continual learning methods, reducing forgetting and favoring faster convergence. Our empirical validation shows that, across datasets and regardless of the number of tasks, our proposed regularization strategy consistently improves baseline performance at the expense of a minimal computational overhead. The lightweight nature of IM ensures that it remains a practical and scalable solution, making it applicable to real-world continual learning scenarios where efficiency is paramount. Finally, we demonstrate the data-agnostic nature of our regularizer by applying it to video data, which presents additional challenges due to its temporal structure and higher memory requirements. Despite the significant domain gap, our experiments show that IM regularizer also improves the performance of video continual learning methods.

Shuming Liu, Chen Zhao, Fatimah Zohra et al.

Temporal action detection (TAD) is a fundamental video understanding task that aims to identify human actions and localize their temporal boundaries in videos. Although this field has achieved remarkable progress in recent years, further progress and real-world applications are impeded by the absence of a standardized framework. Currently, different methods are compared under different implementation settings, evaluation protocols, etc., making it difficult to assess the real effectiveness of a specific technique. To address this issue, we propose \textbf{OpenTAD}, a unified TAD framework consolidating 16 different TAD methods and 9 standard datasets into a modular codebase. In OpenTAD, minimal effort is required to replace one module with a different design, train a feature-based TAD model in end-to-end mode, or switch between the two. OpenTAD also facilitates straightforward benchmarking across various datasets and enables fair and in-depth comparisons among different methods. With OpenTAD, we comprehensively study how innovations in different network components affect detection performance and identify the most effective design choices through extensive experiments. This study has led to a new state-of-the-art TAD method built upon existing techniques for each component. We have made our code and models available at https://github.com/sming256/OpenTAD.

Minseon Kim, Jin Myung Kwak, Lama Alssum et al.

Fine-tuning language models is commonly believed to inevitably harm their safety, i.e., refusing to respond to harmful user requests, even when using harmless datasets, thus requiring additional safety measures. We challenge this belief through systematic testing, showing that poor optimization choices, rather than inherent trade-offs, often cause safety problems, measured as harmful responses to adversarial prompts. By properly selecting key training hyper-parameters, e.g., learning rate, batch size, and gradient steps, we reduce unsafe model responses from 16\% to approximately 5\%, as measured by keyword matching, while maintaining utility performance. Based on this observation, we propose a simple exponential moving average (EMA) momentum technique in parameter space that preserves safety performance by creating a stable optimization path and retains the original pre-trained model's safety properties. Our experiments on the Llama families across multiple datasets (Dolly, Alpaca, ORCA) demonstrate that safety problems during fine-tuning can largely be avoided without specialized interventions, outperforming existing approaches that require additional safety data while offering practical guidelines for maintaining both model performance and safety during adaptation.

Lama Alssum, Juan Leon Alcazar, Merey Ramazanova et al.

Class-incremental learning is one of the most important settings for the study of Continual Learning, as it closely resembles real-world application scenarios. With constrained memory sizes, catastrophic forgetting arises as the number of classes/tasks increases. Studying continual learning in the video domain poses even more challenges, as video data contains a large number of frames, which places a higher burden on the replay memory. The current common practice is to sub-sample frames from the video stream and store them in the replay memory. In this paper, we propose SMILE a novel replay mechanism for effective video continual learning based on individual/single frames. Through extensive experimentation, we show that under extreme memory constraints, video diversity plays a more significant role than temporal information. Therefore, our method focuses on learning from a small number of frames that represent a large number of unique videos. On three representative video datasets, Kinetics, UCF101, and ActivityNet, the proposed method achieves state-of-the-art performance, outperforming the previous state-of-the-art by up to 21.49%.