Operator fusion, as a key performance optimization technique in the deployment of AI models, significantly improves execution efficiency and has been widely adopted in modern AI compilers. However, for cascaded reduction operations involving multiple loops with inter-loop data dependencies, such as the safe softmax followed by GEMM within attention mechanisms, existing compilers lack effective automated fusion and kernel generation capabilities. Although some works have addressed specific instances through hand-crafted fusion strategies, their solutions are limited in generality and difficult to extend to other similar structures. Given the prevalence of such computational patterns in deep learning models, there remains significant untapped potential in achieving general and automated fusion optimization. In this paper, we present a formal theoretical methodology for analyzing cascaded reductions which can fuse them into a single loop and introduce an incremental computation form. Based on this methodology, we design Red uction Fuser (RedFuser), a framework that automatically identifies supported cascaded reduction patterns and generates optimized fused kernels. Experiments show that RedFuser successfully fuses diverse workloads, achieving up to 2× to 5× speedup over state-of-the-art AI compilers and matching the performance of highly optimized hand-written kernels.

As an important and challenging task in video understanding, Temporal Action Detection (TAD) has been deeply studied in recent years. However, current works mainly tackle this task with visual information, while neglecting to explore the potential of the audio modality. To address this challenge, in this paper, we propose a simple yet effective AudioVisual Temporal Action Detection Transformer named AV- TAD, which performs early fusion on audio and visual modalities in an end-to-end fashion. On top of it, a novel query formulation is introduced by directly adopting temporal segment coordinates as queries in Transformer decoder, thus allowing us to perform dynamic segment update layer-by-layer. To the best of our knowledge, this is the first attempt to investigate both audio and video feature with a multi-modal Transformer in TAD task. Extensive experiments on THUMOS14 dataset demonstrate that our proposed AV-TAD can outperform the previous methods by a clear margin.

Deep learning-based methods for human pose estimation require large volumes of training data to achieve superior performance. However, data acquisition in classroom environments raises privacy concerns, which will undoubtedly hinder the development of the latest deep learning techniques in education domain. Due to the absence of large, richly annotated classroom datasets, research into classroom observation has had to be done by manually collecting and annotating datasets. Unfortunately, the annotation of such data is time-consuming and challenging in over-crowded classrooms. To break through these limitations, we open source SynPose, a large, densely labeled synthetic dataset specifically designed for crowded human pose estimation in classroom and meeting scenarios. Moreover, we propose a novel CTGAN to bridge the domain gap. Comprehensive experiments on real-world classroom images show that our proposed dataset and method deliver important performance benefits compared to existing datasets, revealing the potential of SynPose for future studies.