ABSTRACT Background Large Language Models (LLMs) have demonstrated remarkable capabilities across various tasks. However, their substantial computational and memory requirements present significant challenges for widespread deployment on edge devices. Motivation In long‐context scenarios, even sub‐billion parameter LLMs face unavoidable memory and performance bottlenecks due to inefficient KV Cache utilization. Existing quantization methods fail to address these challenges effectively. Method This paper addresses these challenges by introducing advanced quantization techniques tailored for sub‐billion parameter LLMs. It specifically targets reducing memory consumption through the conversion of the model's KV Cache to lower‐bit integers. We present SubKV, a quantization method specifically designed to optimize the KV Cache in sub‐billion parameter LLMs. Our analysis reveals distinct distributional differences in the magnitude of key and value caches. Leveraging this insight, we apply Per‐Channel Quantization to the key cache and Per‐Token Quantization to the value cache. Furthermore, we introduce the Dynamic Window Quantization method to enhance attention computations. To mitigate the extreme sensitivity of the first token, we also introduce Attention Sink‐Aware Quantization. Results Experimental results demonstrate that SubKV significantly reduces the KV Cache size during long context inference while maintaining model performance, offering superior results to existing KV Cache quantization methods.

While vision transformers (ViTs) have continued to achieve new milestones in computer vision, their complicated network architectures with high computation and memory costs have hindered their deployment on resource-limited edge devices. Some customized accelerators have been proposed to accelerate the execution of ViTs, achieving improved performance with reduced energy consumption. However, these approaches utilize flattened attention mechanisms and ignore the inherent hierarchical visual semantics in images. In this work, we conduct a thorough analysis of hierarchical visual semantics in real-world images, revealing opportunities and challenges of leveraging visual semantics to accelerate ViTs. We propose ReViT, a systematic algorithm and architecture co-design approach, which aims to exploit the visual semantics to accelerate ViTs. Our proposed algorithm can leverage the same semantic class with strong feature similarity to reduce computation and communication in a differential attention mechanism, and support the semantic-aware attention efficiently. A novel dedicated architecture is designed to support the proposed algorithm and translate it into performance improvements. Moreover, we propose an efficient execution dataflow to alleviate workload imbalance and maximize hardware utilization. ReViT opens new directions for accelerating ViTs by exploring the underlying visual semantics of images. ReViT gains an average of 2.3$\boldsymbol{\times}$speedup and 3.6$\boldsymbol{\times}$energy efficiency over state-of-the-art ViT accelerators.

Point cloud neural networks are gaining increasing attention in emerging 3-D computer vision applications, such as autonomous driving, robotics, and virtual reality. Many customized accelerators for 3-D point clouds have been developed to pursue superior time and energy efficiencies. In this work, we reveal that spatially adjacent points in a 3-D point cloud show similar feature values and relationships, implying substantial redundant computations and memory accesses, while which have been previously ignored. To reduce such redundancies, we propose SimDiff, an algorithm-accelerator co-design framework that boosts 3-D point cloud processing by cleverly leveraging spatial similarity toward excellent speedup and energy efficiency. On the algorithm side, we design a novel similarity-aware differential point cloud neural network (dubbed SD-PCNet). Differing from the standard flow of mainstream point cloud networks, it abstracts a brand-new execution flow for point cloud processing by utilizing spatial similarity among points and dynamic differential execution. On the accelerator side, we propose SD-PCAcc, a supporting accelerator to convert algorithm-level redundancy reductions into performance enhancements. On the deployment side, we propose efficient strategies for network-to-accelerator mapping and scheduling, high-bandwidth memory (HBM) channel allocation, and core component reconfiguration, facilitating the proposed methodologies into practical implementation. Extensive evaluation results show that, with preserved accuracy, our SimDiff gains an average of$3.2\times $speedup and$3.1\times $energy efficiency compared to the state-of-the-art competitors.

Entity alignment aims to use pre-aligned seed pairs to find other equivalent entities from different knowledge graphs and is widely used in graph fusion-related fields. However, as the scale of knowledge graphs increases, manually annotating pre-aligned seed pairs becomes difficult. Existing research utilizes entity embeddings obtained by aggregating single structural information to identify potential seed pairs, thus reducing the reliance on pre-aligned seed pairs. However, due to the structural heterogeneity of KG, the quality of potential seed pairs obtained using only a single structural information is not ideal. In addition, although existing research improves the quality of potential seed pairs through semi-supervised iteration, they underestimate the impact of embedding distortion produced by noisy seed pairs on the alignment effect. In order to solve the above problems, we propose a seed expanded-aware graph neural network with iterative optimization for semi-supervised entity alignment, named SE-GNN. First, we utilize the semantic attributes and structural features of entities, combined with a conditional filtering mechanism, to obtain high-quality initial potential seed pairs. Next, we designed a local and global awareness mechanism. It introduces initial potential seed pairs and combines local and global information to obtain a more comprehensive entity embedding representation, which alleviates the impact of KG structural heterogeneity and lays the foundation for the optimization of initial potential seed pairs. Then, we designed the threshold nearest neighbor embedding correction strategy. It combines the similarity threshold and the bidirectional nearest neighbor method as a filtering mechanism to select iterative potential seed pairs and also uses an embedding correction strategy to eliminate the embedding distortion. Finally, we will reach the optimized potential seeds after iterative rounds to input local and global sensing mechanisms, obtain the final entity embedding, and perform entity alignment. Experimental results on public datasets demonstrate the excellent performance of our SE-GNN, showcasing the effectiveness of the model. Our code is publicly available athttps://github.com/ShuoShan1/SE-GNN.

Since Multimodal Emotion Recognition in Conversation (MERC) can be applied to public opinion monitoring, intelligent dialogue robots, and other fields, it has received extensive research attention in recent years. Unlike traditional unimodal emotion recognition, MERC can fuse complementary semantic information between multiple modalities (e.g., text, audio, and vision) to improve emotion recognition. However, previous work ignored the inter-modal alignment process and the intra-modal noise information before multimodal fusion but directly fuses multimodal features, which will hinder the model for representation learning. In this study, we have developed a novel approach called Masked Graph Learning with Recursive Alignment (MGLRA) to tackle this problem, which uses a recurrent iterative module with memory to align multimodal features, and then uses the masked GCN for multimodal feature fusion. First, we employ LSTM to capture contextual information and use a graph attention-filtering mechanism to eliminate noise effectively within the modality. Second, we build a recurrent iteration module with a memory function, which can use communication between different modalities to eliminate the gap between modalities and achieve the preliminary alignment of features between modalities. Then, a cross-modal multi-head attention mechanism is introduced to achieve feature alignment between modalities and construct a masked GCN for multimodal feature fusion, which can perform random mask reconstruction on the nodes in the graph to obtain better node feature representation. Finally, we utilize a multilayer perceptron (MLP) for emotion recognition. Extensive experiments on two benchmark datasets (i.e., IEMOCAP and MELD) demonstrate that MGLRA outperforms state-of-the-art methods.

In terms of human-computer interaction, it is becoming more and more important to correctly understand the user’s emotional state in a conversation, so the task of multimodal emotion recognition (MER) started to receive more attention. However, existing emotion classification methods usually perform classification only once. Sentences are likely to be misclassified in a single round of classification. Previous work usually ignores the similarities and differences between different morphological features in the fusion process. To address the above issues, we propose a two-stage emotion recognition model based on graph contrastive learning (TS-GCL). First, we encode the original dataset with different preprocessing modalities. Second, a graph contrastive learning (GCL) strategy is introduced for these three modal data with other structures to learn similarities and differences within and between modalities. Finally, we use MLP twice to achieve the final emotion classification. This staged classification method can help the model to better focus on different levels of emotional information, thereby improving the performance of the model. Extensive experiments show that TS-GCL has superior performance on IEMOCAP and MELD datasets compared with previous methods.

Deep learning on point clouds has attracted increasing attention for various emerging 3D computer vision applications, such as autonomous driving, robotics, and virtual reality. These applications interact with people in real-time on edge devices and thus require low latency and low energy. To accelerate the execution of deep neural networks (DNNs) on point clouds, some customized accelerators have been proposed, which achieved a significantly higher performance with reduced energy consumption than GPUs and existing DNN accelerators.