Momentum-Guided Semantic Forecasting (MoFore) for Self-Supervised Video Representation Learning

Researchers have introduced two new frameworks for advancing self-supervised video representation learning and multi-label abnormality analysis from 3D CT scans, demonstrating improved performance and applicability.

A Momentum-Guided Semantic Forecasting framework (MoFore) has been proposed for self-supervised video representation learning. MoFore learns temporally predictive video representations by forecasting future latent embeddings from temporally distant context clips^[1]. The framework combines predictive latent forecasting with contrastive regularization to encourage temporal consistency. Experiments on the UCF101 dataset show that MoFore learns temporally consistent and semantically meaningful video representations without using action labels during training. Meanwhile, a graph-based framework has been proposed for multi-label abnormality analysis from 3D CT scans. This method represents 3D CT volumes as structured graphs, where axial slice triplets serve as nodes processed through spectral graph convolution^[2]. The model was trained and evaluated on 3 datasets from independent institutions, showing competitive performance compared to state-of-the-art visual encoders. Comprehensive ablation studies were conducted to evaluate various aggregation strategies and graph connectivity patterns.