14 March 2024 in IEEE Transactions on Pattern Analysis and Machine Intelligence
This paper presents a novel spatiotemporal transformer network that introduces several original components to detect actions in untrimmed videos. First, the multi-feature selective semantic attention model calculates the correlations between spatial and motion features to model spatiotemporal interactions between different action semantics properly. Second, the motion-aware network encodes the locations of action semantics in video frames utilizing the motion-aware 2D positional encoding algorithm. Such a motion-aware mechanism memorizes the dynamic spatiotemporal variations in action frames that current methods cannot exploit. Third, the sequence-based temporal attention model captures the heterogeneous temporal dependencies in action frames. In contrast to standard temporal attention used in natural language processing, primarily aimed at finding similarities between linguistic words, the proposed sequence-based temporal attention is designed to determine both the differences and similarities between video frames that jointly define the meaning of actions. The proposed approach outperforms the state-of-the-art solutions on four spatiotemporal action datasets: AVA 2.2, AVA 2.1, UCF101-24, and EPIC-Kitchens.
2 November 2025 in Pattern Recognition
24 October 2025 in IEEE Signal Processing Letters
We present a causal state space model (CSSM) for video understanding that couples a learned causal DAG with latent state dynamics. Latent factors form DAG nodes, enabling explicit cause–effect modeling over time; the state-space form provides efficient sequence inference, while the graph adds interpretability and robustness to distribution shifts. We learn the latent graph and inject its adjacency into the transition operator. On HMDB-51, UCF-101, and HAR, CSSM improves accuracy over strong baselines and supports counterfactual reasoning about video events.
5 November 2026 in npj Artificial Intelligence
Deep networks are usually built first and trained later, forcing fixed capacity and wasted compute. We introduce the Autoplastic Neural Network, a training-time framework that learns structure and weights together. Layer inclusion, neuron-level connectivity, and per-layer operation mixing are jointly optimized within a fixed maximal supernetwork, whose pre-allocated components are reinforced or pruned as learning progresses; the resulting model supports per-input routing and can also be exported as a compact static architecture. A stability-gated rule coordinates architectural changes with gradient descent and yields convergence guarantees. Across vision, video, language, bioinformatics, and healthcare tasks, the approach surpasses strong static and dynamic baselines while using less computation. Unlike neural architecture search, no outer search loop is required; evolution happens inside a single learning process with cost comparable to training a single model. By unifying optimization and design, Autoplastic networks allocate capacity where needed, remove redundancy when possible, and deliver practical, resource-aware deep learning.