A.1 Motivation For what purpose was the dataset created? Who created the dataset (e.g., which team, research group) and on behalf of which entity (e.g., Who funded the creation of the dataset? What do the instances that comprise the dataset represent (e.g., documents, photos, people, How many instances are there in total (of each type, if appropriate)? The SRFUND dataset contains all possible instances. What data does each instance consist of?

We thank the reviewers for their constructive feedback and address their comments below. In this paper, we focus on the models with low memory budgets. Empirically, we also observe that edge probabilities converge to 0 or 1. Y es, in our model edge indicators are independent random variables. Furthermore, PRODIGE is a general method that works for a variety of tasks (e.g. If accepted, we will include a more detailed comparison of the two methods with explanation.

ViL T for each task, and details about how low-shot versions of each task are sampled. B.1 Applying ViL T to Multi-Choice T asks B.1.1 Applying ViL T to VCR We follow previous work [Zellers et al., 2021, Hessel et al., 2022] and draw colored boxes directly The grounded text references, e.g. We follow the original implementations [Zellers et al., 2019b, Bisk et al., 2020] to model these tasks, B.2 Applying ViL T to Unimodal T asks We conduct low-shot experiments to test the model's transferability to unimodal However, different sub-samples the training set may lead to different results. For vision-only tasks, we found that simply using "This is an image." We also conduct ablation studies that include two baselines: (1) not inputting any image to ViL T at all, and (2) inputting the zero-vector image instead of the average image of the COCO dataset.

This paper presents our system for the SMM4H-HeaRD 2025 shared tasks, specifically Task 4 (Subtasks 1, 2a, and 2b) and Task 5 (Subtasks 1 and 2). Task 4 focused on detecting mentions of insomnia in clinical notes, while Task 5 addressed the extraction of food safety events from news articles. We participated in all subtasks and report key findings across them, with particular emphasis on Task 5 Subtask 1, where our system achieved strong performance--securing first place with an F1 score of 0.958 on the test set. To attain this result, we employed encoder-based models (e.g., RoBERTa), alongside GPT -4 for data augmentation. This paper outlines our approach, including preprocessing, model architecture, and subtask-specific adaptations.

A.1 PAC Bayesian Bound In this part, we provide a detailed PAC-Bound based on the continual learning scenario. Given a "prior" distribution P (a common assumption is zero mean, σ We now consider the bound in the continual learning scenario. Based on Eq. (6), the expected error of f Note that we only consider one gradient update to v in the second equation for simplicity, but using multiple gradient updates is a straightforward extension. The importance of each basis is constrained to be between 0 and 1, where 0 indicates that the basis is not important to old tasks and can completely release for learning new tasks. Similar to [34], we calculate the bases of these subspaces for each layer by analyzing network representations after learning each task with Singular Value Decomposition (SVD), and then use it to update v and w by layer.

Continual learning in deep neural networks often suffers from catastrophic forgetting, where representations for previous tasks are overwritten during subsequent training. We propose a novel sample retrieval strategy from the memory buffer that leverages both gradient-conflicting and gradient-aligned samples to effectively retain knowledge about past tasks within a supervised contrastive learning framework. Gradient-conflicting samples are selected for their potential to reduce interference by re-aligning gradients, thereby preserving past task knowledge. Meanwhile, gradient-aligned samples are incorporated to reinforce stable, shared representations across tasks. By balancing gradient correction from conflicting samples with alignment reinforcement from aligned ones, our approach increases the diversity among retrieved instances and achieves superior alignment in parameter space, significantly enhancing knowledge retention and mitigating proxy drift. Empirical results demonstrate that using both sample types outperforms methods relying solely on one sample type or random retrieval. Experiments on popular continual learning benchmarks in computer vision validate our method's state-of-the-art performance in mitigating forgetting while maintaining competitive accuracy on new tasks.

Deep learning systems have become increasingly energy- and computation-intensive, raising concerns about their environmental impact. As organizers of the Detection and Classification of Acoustic Scenes and Events (DCASE) challenge, we recognize the importance of addressing this issue. For the past three years, we have integrated energy consumption metrics into the evaluation of sound event detection (SED) systems. In this paper, we analyze the impact of this energy criterion on the challenge results and explore the evolution of system complexity and energy consumption over the years. We highlight a shift towards more energy-efficient approaches during training without compromising performance, while the number of operations and system complexity continue to grow. Through this analysis, we hope to promote more environmentally friendly practices within the SED community.