Generalizable Video Quality Assessment via Weak-to-Strong Learning

A new machine learning framework could reshape how video quality is assessed by eliminating the need for human-labeled training data. Researchers propose a weak-to-strong learning paradigm that allows a model to surpass its teacher, achieving state-of-the-art results on both standard and out-of-distribution benchmarks [1]. The dominant approach to video quality assessment (VQA) relies on supervised training with datasets labeled by human viewers. Despite progress, these models often fail to generalize to video content they have not seen before [1]. The new work, detailed in a paper by Linhan Cao and colleagues, explores weak-to-strong (W2S) learning as an alternative that does not depend on human annotations [1]. Machine learning, broadly, involves developing statistical algorithms that learn from data and generalize to unseen examples, with deep neural networks now powering many advanced applications [3]. The researchers first demonstrated that a straightforward W2S strategy produces a “weak-to-strong effect” in VQA. A strong student model not only matched its weaker teacher on familiar benchmarks but outperformed it on out-of-distribution tests [1]. Neural networks, which consist of layers of interconnected artificial neurons that adjust connection weights during training, underpin such student-teacher architectures [4]. Building on this finding, the team designed a framework that integrates supervision signals from multiple VQA teachers, including existing off-the-shelf models and synthetic distortion simulators. These signals are combined through a learn-to-rank formulation [1]. The framework also employs iterative W2S training: after each cycle, the strong student becomes the teacher for the next round, progressively concentrating on more difficult cases [1]. Extensive experiments showed the method achieved state-of-the-art performance, with particularly significant gains in out-of-distribution scenarios [1]. The authors argue that W2S learning offers a principled path to overcoming annotation barriers and achieving scalable generalization in video quality assessment [1]. The paper’s code and data are slated for public release [1].