Qiang Wu

Yuzhang Shang, Zhihang Yuan, Qiang Wu et al. · berkeley

This paper explores network binarization, a radical form of quantization, compressing model weights to a single bit, specifically for Large Language Models (LLMs) compression. Due to previous binarization methods collapsing LLMs, we propose a novel approach, Partially-Binarized LLM (PB-LLM), which can achieve extreme low-bit quantization while maintaining the linguistic reasoning capacity of quantized LLMs. Specifically, our exploration first uncovers the ineffectiveness of naive applications of existing binarization algorithms and highlights the imperative role of salient weights in achieving low-bit quantization. Thus, PB-LLM filters a small ratio of salient weights during binarization, allocating them to higher-bit storage, i.e., partially-binarization. PB-LLM is extended to recover the capacities of quantized LMMs, by analyzing from the perspective of post-training quantization (PTQ) and quantization-aware training (QAT). Under PTQ, combining the concepts from GPTQ, we reconstruct the binarized weight matrix guided by the Hessian matrix and successfully recover the reasoning capacity of PB-LLM in low-bit. Under QAT, we freeze the salient weights during training, explore the derivation of optimal scaling factors crucial for minimizing the quantization error, and propose a scaling mechanism based on this derived scaling strategy for residual binarized weights. Those explorations and the developed methodologies significantly contribute to rejuvenating the performance of low-bit quantized LLMs and present substantial advancements in the field of network binarization for LLMs.The code is available at https://github.com/hahnyuan/BinaryLLM.

Zhihang Yuan, Lin Niu, Jiawei Liu et al.

Large-scale language models (LLMs) have demonstrated impressive performance, but their deployment presents challenges due to their significant memory usage. This issue can be alleviated through quantization. In this paper, we identify that the challenge in quantizing activations in LLMs arises from varying ranges across channels, rather than solely the presence of outliers. To address this challenge, we introduce a quantization method called RPTQ, which utilizes a reorder-based approach. By rearranging the channels and quantizing them in clusters, RPTQ effectively mitigates the impact of range differences between channels. To minimize the overhead of the reorder operation, we fuse it into the layer norm operation and weights in linear layers. In our experiments, RPTQ achieved a significant breakthrough by utilizing 3-bit activation in LLMs for the first time, resulting in a substantial reduction in memory usage. For instance, quantizing OPT-175b can lead to a memory consumption reduction of up to 80%.

Yunxiao Shi, Xing Zi, Zijing Shi et al.

Retrieval-augmented generation (RAG) techniques leverage the in-context learning capabilities of large language models (LLMs) to produce more accurate and relevant responses. Originating from the simple 'retrieve-then-read' approach, the RAG framework has evolved into a highly flexible and modular paradigm. A critical component, the Query Rewriter module, enhances knowledge retrieval by generating a search-friendly query. This method aligns input questions more closely with the knowledge base. Our research identifies opportunities to enhance the Query Rewriter module to Query Rewriter+ by generating multiple queries to overcome the Information Plateaus associated with a single query and by rewriting questions to eliminate Ambiguity, thereby clarifying the underlying intent. We also find that current RAG systems exhibit issues with Irrelevant Knowledge; to overcome this, we propose the Knowledge Filter. These two modules are both based on the instruction-tuned Gemma-2B model, which together enhance response quality. The final identified issue is Redundant Retrieval; we introduce the Memory Knowledge Reservoir and the Retriever Trigger to solve this. The former supports the dynamic expansion of the RAG system's knowledge base in a parameter-free manner, while the latter optimizes the cost for accessing external knowledge, thereby improving resource utilization and response efficiency. These four RAG modules synergistically improve the response quality and efficiency of the RAG system. The effectiveness of these modules has been validated through experiments and ablation studies across six common QA datasets. The source code can be accessed at https://github.com/Ancientshi/ERM4.

Zhihang Yuan, Jiawei Liu, Jiaxiang Wu et al.

Post-training quantization (PTQ) is a popular method for compressing deep neural networks (DNNs) without modifying their original architecture or training procedures. Despite its effectiveness and convenience, the reliability of PTQ methods in the presence of some extrem cases such as distribution shift and data noise remains largely unexplored. This paper first investigates this problem on various commonly-used PTQ methods. We aim to answer several research questions related to the influence of calibration set distribution variations, calibration paradigm selection, and data augmentation or sampling strategies on PTQ reliability. A systematic evaluation process is conducted across a wide range of tasks and commonly-used PTQ paradigms. The results show that most existing PTQ methods are not reliable enough in term of the worst-case group performance, highlighting the need for more robust methods. Our findings provide insights for developing PTQ methods that can effectively handle distribution shift scenarios and enable the deployment of quantized DNNs in real-world applications.

Bo Qian, Hao Chen, Xiangning Wang et al.

Computer-assisted automatic analysis of diabetic retinopathy (DR) is of great importance in reducing the risks of vision loss and even blindness. Ultra-wide optical coherence tomography angiography (UW-OCTA) is a non-invasive and safe imaging modality in DR diagnosis system, but there is a lack of publicly available benchmarks for model development and evaluation. To promote further research and scientific benchmarking for diabetic retinopathy analysis using UW-OCTA images, we organized a challenge named "DRAC - Diabetic Retinopathy Analysis Challenge" in conjunction with the 25th International Conference on Medical Image Computing and Computer Assisted Intervention (MICCAI 2022). The challenge consists of three tasks: segmentation of DR lesions, image quality assessment and DR grading. The scientific community responded positively to the challenge, with 11, 12, and 13 teams from geographically diverse institutes submitting different solutions in these three tasks, respectively. This paper presents a summary and analysis of the top-performing solutions and results for each task of the challenge. The obtained results from top algorithms indicate the importance of data augmentation, model architecture and ensemble of networks in improving the performance of deep learning models. These findings have the potential to enable new developments in diabetic retinopathy analysis. The challenge remains open for post-challenge registrations and submissions for benchmarking future methodology developments.

Yunxiao Shi, Wujiang Xu, Haimin Zhang et al.

Modeling high-order feature interactions is crucial for click-through rate (CTR) prediction, yet traditional approaches typically predefine a maximum interaction order and exhaustively enumerate feature combinations up to that order. This paradigm depends heavily on prior domain knowledge to delimit the interaction space and incurs substantial computational overhead. As a result, conventional CTR models face a persistent tension between enriching representations with complex high-order interactions and keeping computation tractable. To address this dual challenge, this study introduces the Kolmogorov-Arnold Represented Sparse Efficient Interaction Network (KarSein). Drawing inspiration from the learnable activation mechanism in the Kolmogorov-Arnold Network (KAN), KarSein leverages this mechanism to adaptively transform low-order basic features into high-order feature interactions, offering a novel approach to feature interaction modeling. KarSein extends the capabilities of KAN by introducing a more efficient architecture that significantly reduces computational costs while accommodating two-dimensional embedding vectors as feature inputs. Furthermore, it overcomes the limitation of KAN's its inability to spontaneously capture multiplicative relationships among features. Extensive experiments highlight the superiority of KarSein, demonstrating its ability to surpass not only the vanilla implementation of KAN in CTR prediction tasks but also other baseline methods. Remarkably, KarSein achieves exceptional predictive accuracy while maintaining a highly compact parameter size and minimal computational overhead. Moreover, KarSein exhibits strong interpretability and structural sparsity. As the first systematic adaptation of KAN to CTR prediction, KarSein offers a practical, parameter-efficient, and interpretable alternative for modeling complex feature interactions in CTR prediction.

Youpeng Zhao, Huadong Tang, Yingying Jiang et al.

Recent advances in vision transformers (ViTs) have achieved great performance in visual recognition tasks. Convolutional neural networks (CNNs) exploit spatial inductive bias to learn visual representations, but these networks are spatially local. ViTs can learn global representations with their self-attention mechanism, but they are usually heavy-weight and unsuitable for mobile devices. In this paper, we propose cross feature attention (XFA) to bring down computation cost for transformers, and combine efficient mobile CNNs to form a novel efficient light-weight CNN-ViT hybrid model, XFormer, which can serve as a general-purpose backbone to learn both global and local representation. Experimental results show that XFormer outperforms numerous CNN and ViT-based models across different tasks and datasets. On ImageNet1K dataset, XFormer achieves top-1 accuracy of 78.5% with 5.5 million parameters, which is 2.2% and 6.3% more accurate than EfficientNet-B0 (CNN-based) and DeiT (ViT-based) for similar number of parameters. Our model also performs well when transferring to object detection and semantic segmentation tasks. On MS COCO dataset, XFormer exceeds MobileNetV2 by 10.5 AP (22.7 -> 33.2 AP) in YOLOv3 framework with only 6.3M parameters and 3.8G FLOPs. On Cityscapes dataset, with only a simple all-MLP decoder, XFormer achieves mIoU of 78.5 and FPS of 15.3, surpassing state-of-the-art lightweight segmentation networks.

Litao Yu, Zhibin Li, Jian Zhang et al.

Scene segmentation in images is a fundamental yet challenging problem in visual content understanding, which is to learn a model to assign every image pixel to a categorical label. One of the challenges for this learning task is to consider the spatial and semantic relationships to obtain descriptive feature representations, so learning the feature maps from multiple scales is a common practice in scene segmentation. In this paper, we explore the effective use of self-attention within multi-scale image windows to learn descriptive visual features, then propose three different strategies to aggregate these feature maps to decode the feature representation for dense prediction. Our design is based on the recently proposed Swin Transformer models, which totally discards convolution operations. With the simple yet effective multi-scale feature learning and aggregation, our models achieve very promising performance on four public scene segmentation datasets, PASCAL VOC2012, COCO-Stuff 10K, ADE20K and Cityscapes.

Sicheng Yang, Xing Hu, Qiang Wu et al.

Vector quantization (VQ) transforms continuous image features into discrete representations, providing compressed, tokenized inputs for generative models. However, VQ-based frameworks suffer from several issues, such as non-smooth latent spaces, weak alignment between representations before and after quantization, and poor coherence between the continuous and discrete domains. These issues lead to unstable codeword learning and underutilized codebooks, ultimately degrading the performance of both reconstruction and downstream generation tasks. To this end, we propose VAEVQ, which comprises three key components: (1) Variational Latent Quantization (VLQ), replacing the AE with a VAE for quantization to leverage its structured and smooth latent space, thereby facilitating more effective codeword activation; (2) Representation Coherence Strategy (RCS), adaptively modulating the alignment strength between pre- and post-quantization features to enhance consistency and prevent overfitting to noise; and (3) Distribution Consistency Regularization (DCR), aligning the entire codebook distribution with the continuous latent distribution to improve utilization. Extensive experiments on two benchmark datasets demonstrate that VAEVQ outperforms state-of-the-art methods.

Wenbo Xu, Yanan Wu, Haoran Jiang et al.

Incremental Few-Shot Semantic Segmentation (iFSS) tackles a task that requires a model to continually expand its segmentation capability on novel classes using only a few annotated examples. Typical incremental approaches encounter a challenge that the objective of the base training phase (fitting base classes with sufficient instances) does not align with the incremental learning phase (rapidly adapting to new classes with less forgetting). This disconnect can result in suboptimal performance in the incremental setting. This study introduces a meta-learning-based prototype approach that encourages the model to learn how to adapt quickly while preserving previous knowledge. Concretely, we mimic the incremental evaluation protocol during the base training session by sampling a sequence of pseudo-incremental tasks. Each task in the simulated sequence is trained using a meta-objective to enable rapid adaptation without forgetting. To enhance discrimination among class prototypes, we introduce prototype space redistribution learning, which dynamically updates class prototypes to establish optimal inter-prototype boundaries within the prototype space. Extensive experiments on iFSS datasets built upon PASCAL and COCO benchmarks show the advanced performance of the proposed approach, offering valuable insights for addressing iFSS challenges.

Yongze Wang, Haimin Zhang, Qiang Wu et al.

Graph neural networks (GNNs) have shown great success in learning from graph-based data. The key mechanism of current GNNs is message passing, where a node's feature is updated based on the information passing from its local neighbourhood. A limitation of this mechanism is that node features become increasingly dominated by the information aggregated from the neighbourhood as we use more rounds of message passing. Consequently, as the GNN layers become deeper, adjacent node features tends to be similar, making it more difficult for GNNs to distinguish adjacent nodes, thereby, limiting the performance of GNNs. In this paper, we propose conditional local feature encoding (CLFE) to help prevent the problem of node features being dominated by the information from local neighbourhood. The idea of our method is to extract the node hidden state embedding from message passing process and concatenate it with the nodes feature from previous stage, then we utilise linear transformation to form a CLFE based on the concatenated vector. The CLFE will form the layer output to better preserve node-specific information, thus help to improve the performance of GNN models. To verify the feasibility of our method, we conducted extensive experiments on seven benchmark datasets for four graph domain tasks: super-pixel graph classification, node classification, link prediction, and graph regression. The experimental results consistently demonstrate that our method improves model performance across a variety of baseline GNN models for all four tasks.

Zukang Xu, Xing Hu, Qiang Wu et al.

Large language models (LLMs) have demonstrated remarkable performance across a wide range of natural language processing tasks. However, their exponentially increasing parameters pose significant challenges for deployment on resource-constrained devices. Vector Quantization (VQ) shows great promise for low-bit quantization (e.g., 2 to 4 bits), but existing work faces two key challenges: unconstrained direction error and suboptimal bit allocation. In this paper, we propose RSAVQ, a novel VQ framework to enhance extremely low-bit quantization for LLMs. RSAVQ introduces two geometry-driven innovations that effectively mitigate above limitations: (1) Error Direction Sensitivity Guidance (EDSG), which leverages the Fisher Information Matrix (FIM)-induced Riemannian metric to project quantization errors onto low-sensitivity directions in the parameter space. Specifically, this projection is performed along the negative natural gradient direction, which effectively suppresses error expansion. (2) Weight Channel Sensitivity Guidance (WCSG) , which constructs a channel-wise sensitivity metric via FIM curvature analysis to dynamically guide bit resource allocation. The approach facilitates a globally optimal quantization solution within prescribed bit constraints. Experiments demonstrate that RSAVQ outperforms existing methods for LLMs. For example, in 2-bit quantization of LLaMA-3 8B, RSAVQ leads baselines like VPTQ and QuIP# by 0.4 in perplexity (PPL) and 1.5 in zero-shot accuracy. This work offers a practical solution for constrained environments and a theoretical bridge between information geometry and the quantization of neural networks, advancing efficient deep learning.

Yunxiao Shi, Shuo Yang, Haimin Zhang et al.

Neural Collaborative Filtering models are widely used in recommender systems but are typically trained under static settings, assuming fixed data distributions. This limits their applicability in dynamic environments where user preferences evolve. Incremental learning offers a promising solution, yet conventional methods from computer vision or NLP face challenges in recommendation tasks due to data sparsity and distinct task paradigms. Existing approaches for neural recommenders remain limited and often lack generalizability. To address this, we propose MEGG, Replay Samples with Maximally Extreme GGscore, an experience replay based incremental learning framework. MEGG introduces GGscore, a novel metric that quantifies sample influence, enabling the selective replay of highly influential samples to mitigate catastrophic forgetting. Being model-agnostic, MEGG integrates seamlessly across architectures and frameworks. Experiments on three neural models and four benchmark datasets show superior performance over state-of-the-art baselines, with strong scalability, efficiency, and robustness. Implementation will be released publicly upon acceptance.

Yunxiao Shi, Xing Zi, Zijing Shi et al.

Retrieval-augmented generation (RAG) for language models significantly improves language understanding systems. The basic retrieval-then-read pipeline of response generation has evolved into a more extended process due to the integration of various components, sometimes even forming loop structures. Despite its advancements in improving response accuracy, challenges like poor retrieval quality for complex questions that require the search of multifaceted semantic information, inefficiencies in knowledge re-retrieval during long-term serving, and lack of personalized responses persist. Motivated by transcending these limitations, we introduce ERAGent, a cutting-edge framework that embodies an advancement in the RAG area. Our contribution is the introduction of the synergistically operated module: Enhanced Question Rewriter and Knowledge Filter, for better retrieval quality. Retrieval Trigger is incorporated to curtail extraneous external knowledge retrieval without sacrificing response quality. ERAGent also personalizes responses by incorporating a learned user profile. The efficiency and personalization characteristics of ERAGent are supported by the Experiential Learner module which makes the AI assistant being capable of expanding its knowledge and modeling user profile incrementally. Rigorous evaluations across six datasets and three question-answering tasks prove ERAGent's superior accuracy, efficiency, and personalization, emphasizing its potential to advance the RAG field and its applicability in practical systems.

Zhihang Yuan, Yuzhang Shang, Yue Song et al.

In this paper, we introduce a new post-training compression paradigm for Large Language Models (LLMs) to facilitate their wider adoption. We delve into LLM weight low-rank decomposition, and find that the challenges of this task stem from (1) the distribution variance in the LLM activations and (2) the sensitivity difference among various kinds of layers. To address these issues, we propose a training-free approach called Activation-aware Singular Value Decomposition (ASVD). Specifically, ASVD manages activation outliers by transforming the weight matrix based on the activation distribution. This transformation allows the outliers in the activation matrix to be absorbed into the transformed weight matrix, thereby enhancing decomposition accuracy. Additionally, we propose an efficient iterative calibration process to optimize layer-specific decomposition by addressing the varying sensitivity of different LLM layers. In this way, ASVD can compress a network by 10%-30%. Based on the success of the low-rank decomposition of projection matrices in the self-attention module, we further introduce ASVD to compress the KV cache. By reducing the channel dimension of KV activations, memory requirements for KV cache can be largely reduced. ASVD can further achieve 50% KV cache reductions without performance drop in a training-free manner.

Zhihang Yuan, Chenhao Xue, Yiqi Chen et al.

Quantization is one of the most effective methods to compress neural networks, which has achieved great success on convolutional neural networks (CNNs). Recently, vision transformers have demonstrated great potential in computer vision. However, previous post-training quantization methods performed not well on vision transformer, resulting in more than 1% accuracy drop even in 8-bit quantization. Therefore, we analyze the problems of quantization on vision transformers. We observe the distributions of activation values after softmax and GELU functions are quite different from the Gaussian distribution. We also observe that common quantization metrics, such as MSE and cosine distance, are inaccurate to determine the optimal scaling factor. In this paper, we propose the twin uniform quantization method to reduce the quantization error on these activation values. And we propose to use a Hessian guided metric to evaluate different scaling factors, which improves the accuracy of calibration at a small cost. To enable the fast quantization of vision transformers, we develop an efficient framework, PTQ4ViT. Experiments show the quantized vision transformers achieve near-lossless prediction accuracy (less than 0.5% drop at 8-bit quantization) on the ImageNet classification task.