The frequent retrieval of Key-Value (KV) cache data has emerged as a significant factor contributing to the inefficiency of the inference process in large language models. Previous research has demonstrated that a small subset of critical KV cache tokens largely influences attention outcomes, leading to methods that either employ fixed sparsity patterns or dynamically select critical tokens based on the query. While dynamic sparse patterns have proven to be more effective, they introduce significant computational overhead, as critical tokens must be reselected for each self-attention computation. In this paper, we reveal substantial similarities in KV cache token criticality across neighboring queries, layers, and heads. Motivated by this insight, we propose HShare, a hierarchical KV sharing framework. HShare facilitates the sharing of critical KV cache token indices across layers, heads, and queries, which significantly reduces the computational overhead associated with query-aware dynamic token sparsity. In addition, we introduce a greedy algorithm that dynamically determines the optimal layer-level and head-level sharing configuration for the decoding phase. We evaluate the effectiveness and efficiency of HShare across various tasks using three models: LLaMA2-7b, LLaMA3-70b, and Mistral-7b. Experimental results demonstrate that HShare achieves competitive accuracy with different sharing ratios, while delivering up to an $8.6\times$ speedup in self-attention operations and a $2.7\times$ improvement in end-to-end throughput compared with FlashAttention2 and GPT-fast respectively. The source code is publicly available at ~\url{https://github.com/wuhuaijin/HShare}.

Video-based Autism Spectrum Disorder (ASD) detection is a challenge to most video classification networks due to the high degree of similarity between categories. Bilinear pooling is a second-order method, which is widely used in fine-grained visual recognition. However, the average summation in bilinear pooling limits its ability to perceive spatial information, which is detrimental to fine-grained visual recognition. In this paper, we propose spatial attentional bilinear pooling to enhance its spatial information extraction without significantly increasing the parameters. Further, we propose a fine-grained action recognition network named SA-B3D with LSTM model for video-based ASD detection. The proposed model can focus on more discriminative regions dynamically and effectively. Compared with state-of-the-art models, the proposed model achieves significant improvement on video-based ASD dataset.

The Perimeter Intrusion Detection System (PIDS) has been widely used in many fields since the development of optic-fiber interferometers and intrusion signal recognition models. However, common signal recognition models, such as Support Vector Machines (SVM) and Back Propagation Neural Networks (BPNN), do not perform well in classifying fiber intrusion signals due to the diversity of intrusion signals and the sensitivity of the fiber. In this paper, an optic-fiber based perimeter intrusion detection and recognition system that uses Sagnac interferometers and the optimized curve fitting model is proposed. Experiments on real perimeter intrusions are performed. Comparisons are carried out among our model and the SVM, BPNN models, which prove that our model is more accurate and robust.