Multimodal representation learning combines complementary signals to capture shared semantic information. For dynamic scenes, these representations need to remain consistent across views while preserving distinctions between temporal states. In visual speech representation learning, view variation changes the visibility and reliability of articulation cues, making it difficult to learn such representations from synchronized audio and video. To address this challenge, we propose MuViRL, a self-supervised framework for multimodal view-invariant visual speech representation learning via calibrated view aggregation. The framework combines synchronized 3D Gaussian view synthesis with multimodal self-supervision to learn consistent representations from real and synthesized data. A multimodal decision model learns relative view utility from teacher prediction losses, and its calibrated scores guide the aggregation of complementary information. Aggregation-to-Single Distillation transfers the aggregated predictions to a single branch. Real-synthetic alignment matches synchronized windows across views while preserving their temporal relations. Experiments on real and synthesized multiview data demonstrate that MuViRL improves multimodal representation consistency and visual speech recognition performance across views.
Ellen Su, Andres Potapczynski, Shikai Qiu, Edward Hughes, Andrew Gordon Wilson
原文摘要
Modern systems are increasingly expected to transfer across tasks not specified during training. What data facilitates generalization to unanticipated settings? One hypothesis is that data with more structural information could contain shared circuits and subprograms that may be recycled in a wider array of downstream settings. Epiplexity, a recently proposed measure of the structural information a compute-bounded learner can extract from data, offers a mechanism to reason about this relationship. As existing estimators for epiplexity only operate on data already seen during training, we show how to operationalize epiplexity as an online training signal by greedily maximizing its stepwise gain in two settings. For data selection, we fit scaling laws to the training loss curves of natural data domains to predict the expected epiplexity gain as a function of training tokens, and use this signal to adaptively determine the sampling weights over domains during training. For synthetic data generation, where no domain structure exists, we reward a generator with the change in epiplexity of a learner's synthetic training data and use REINFORCE policy gradients to guide the generator toward an epiplexity-maximizing distribution. In both cases, we find higher epiplexity predicts improved downstream performance on zero-shot and fine-tuning tasks, supporting the hypothesis that data rich in structural information yield representations that better transfer across domains.
Aditya Gupta, Jean Sebastien Dandurand, Kai Qiu, Rohan Choudhury, Laszlo A. Jeni
原文摘要
Vision Transformers (ViTs) are bottlenecked by the length of their input token sequence. Standard patchification fixes this length with a single patch size applied across the image, regardless of regional content. Prior work on adaptive patch sizing reduces token count by assigning coarser patches to visually simple regions, but relies on noisy pixel-level heuristics such as local entropy to make patchification decisions. We argue that this compute allocation should be grounded in the model's interpretation of task relevant information, not in raw pixel statistics. We introduce Split Policy Learned Image Tokenization which learns to allocate token resolution based on what the model finds meaningful, producing shorter token sequences that are better matched to task-relevant content. SPLIT matches the accuracy of full-resolution ViT classification baselines while using 30% fewer tokens. The same gains hold on dense prediction: SPLIT achieves competitive accuracy on COCO and ADE20K, demonstrating that semantically aligned patchification generalises well beyond classification.