To address the communication bottleneck challenge in distributed learning, our work introduces a novel two-stage quantization strategy designed to enhance the communication efficiency of distributed Stochastic Gradient Descent (SGD). The proposed method initially employs truncation to mitigate the impact of long-tail noise, followed by a non-uniform quantization of the post-truncation gradients based on their statistical characteristics. We provide a comprehensive convergence analysis of the quantized distributed SGD, establishing theoretical guarantees for its performance. Furthermore, by minimizing the convergence error, we derive optimal closed-form solutions for the truncation threshold and non-uniform quantization levels under given communication constraints. Both theoretical insights and extensive experimental evaluations demonstrate that our proposed algorithm outperforms existing quantization schemes, striking a superior balance between communication efficiency and convergence performance.
Federated Learning (FL) offers a compelling method for training machine learning models with a focus on preserving data privacy. The presence of system heterogeneity and statistical heterogeneity, recognized challenges in FL, arises from the diversity of client hardware, network, and dataset distribution. This diversity can critically affect the training pace and the performance of models. While many studies address either system or statistical heterogeneity by introducing communication-efficient or stable convergence algorithms, addressing these challenges in isolation often leads to compromises due to unaddressed heterogeneity. In response, this paper introduces FedShift, a novel algorithm designed to enhance both the training speed and the models' accuracy in a dual heterogeneity scenario. Our solution can improve client engagement through quantization and mitigate the adverse effects on performance typically associated with quantization by employing a shifting technique. This technique has proven to enhance accuracy by an average of 3.9% in diverse heterogeneity environments.
The Language of Thought Hypothesis suggests that human cognition operates on a structured, language-like system of mental representations. While neural language models can naturally benefit from the compositional structure inherently and explicitly expressed in language data, learning such representations from non-linguistic general observations, like images, remains a challenge. In this work, we introduce the Neural Language of Thought Model (NLoTM), a novel approach for unsupervised learning of LoTH-inspired representation and generation. NLoTM comprises two key components: (1) the Semantic Vector-Quantized Variational Autoencoder, which learns hierarchical, composable discrete representations aligned with objects and their properties, and (2) the Autoregressive LoT Prior, an autoregressive transformer that learns to generate semantic concept tokens compositionally, capturing the underlying data distribution. We evaluate NLoTM on several 2D and 3D image datasets, demonstrating superior performance in downstream tasks, out-of-distribution generalization, and image generation quality compared to patch-based VQ-VAE and continuous object-centric representations. Our work presents a significant step towards creating neural networks exhibiting more human-like understanding by developing LoT-like representations and offers insights into the intersection of cognitive science and machine learning.
SWIN transformer is a prominent vision transformer model that has state-of-the-art accuracy in image classification tasks. Despite this success, its unique architecture causes slower inference compared with similar deep neural networks. Integer quantization of the model is one of the methods used to improve its inference latency. However, state-of-the-art has not been able to fully quantize the model. In this work, we improve upon the inference latency of the state-of-the-art methods by removing the floating-point operations, which are associated with the GELU activation in Swin Transformer. While previous work proposed to replace the non-integer operations with linear approximation functions, we propose to replace GELU with ReLU activation. The advantage of ReLU over previous methods is its low memory and computation complexity. We use iterative knowledge distillation to compensate for the lost accuracy due to replacing GELU with ReLU. We quantize our GELU-less SWIN transformer and show that on an RTX 4090 NVIDIA GPU we can improve the inference latency of the quantized SWIN transformer by at least $11\%$ while maintaining an accuracy drop of under $0.5\%$ on the ImageNet evaluation dataset.
Manual crowd counting for real-world problems is impossible or results in wildly inaccurate estimations. Deep learning is one area that has been applied to address this issue. Crowd counting is a computationally intensive task. Therefore, many crowd counting models employ large-scale deep convolutional neural networks (CNN) to achieve higher accuracy. However, these are typically at the cost of performance and inference speed. This makes such approaches difficult to apply in real-world settings, e.g., on Internet-of-Things (IoT) devices. To tackle this problem, one method is to compress models using pruning and quantization or use of lightweight model backbones. However, such methods often result in a significant loss of accuracy. To address this, some studies have explored knowledge distillation methods to extract useful information from large state-of-the-art (teacher) models to guide/train smaller (student) models. However, knowledge distillation methods suffer from the problem of information loss caused by hint-transformers. Furthermore, teacher models may have a negative impact on student models. In this work, we propose a method based on knowledge distillation that uses self-transformed hints and loss functions that ignore outliers to tackle real-world and challenging crowd counting tasks. Based on our approach, we achieve a MAE of 77.24 and a MSE of 276.17 using the JHU-CROWD++ [1] test set. This is comparable to state-of-the-art deep crowd counting models, but at a fraction of the original model size and complexity, thus making the solution suitable for IoT devices. The source code is available at https://github.com/huangzuo/effcc_distilled.
Understanding context is key to understanding human language, an ability which Large Language Models (LLMs) have been increasingly seen to demonstrate to an impressive extent. However, though the evaluation of LLMs encompasses various domains within the realm of Natural Language Processing, limited attention has been paid to probing their linguistic capability of understanding contextual features. This paper introduces a context understanding benchmark by adapting existing datasets to suit the evaluation of generative models. This benchmark comprises of four distinct tasks and nine datasets, all featuring prompts designed to assess the models' ability to understand context. First, we evaluate the performance of LLMs under the in-context learning pretraining scenario. Experimental results indicate that pre-trained dense models struggle with understanding more nuanced contextual features when compared to state-of-the-art fine-tuned models. Second, as LLM compression holds growing significance in both research and real-world applications, we assess the context understanding of quantized models under in-context-learning settings. We find that 3-bit post-training quantization leads to varying degrees of performance reduction on our benchmark. We conduct an extensive analysis of these scenarios to substantiate our experimental results.
Over-the-air (OTA) computation has recently emerged as a communication-efficient Federated Learning (FL) paradigm to train machine learning models over wireless networks. However, its performance is limited by the device with the worst SNR, resulting in fast yet noisy updates. On the other hand, allocating orthogonal resource blocks (RB) to individual devices via digital channels mitigates the noise problem, at the cost of increased communication latency. In this paper, we address this discrepancy and present ADFL, a novel Analog-Digital FL scheme: in each round, the parameter server (PS) schedules each device to either upload its gradient via the analog OTA scheme or transmit its quantized gradient over an orthogonal RB using the ``digital" scheme. Focusing on a single FL round, we cast the optimal scheduling problem as the minimization of the mean squared error (MSE) on the estimated global gradient at the PS, subject to a delay constraint, yielding the optimal device scheduling configuration and quantization bits for the digital devices. Our simulation results show that ADFL, by scheduling most of the devices in the OTA scheme while also occasionally employing the digital scheme for a few devices, consistently outperforms OTA-only and digital-only schemes, in both i.i.d. and non-i.i.d. settings.
LLMs are seeing growing use for applications which require large context windows, and with these large context windows KV cache activations surface as the dominant contributor to memory consumption during inference. Quantization is a promising approach for compressing KV cache activations; however, existing solutions fail to represent activations accurately in sub-4-bit precision. Our work, KVQuant, facilitates low precision KV cache quantization by incorporating several novel methods: (i) Per-Channel Key Quantization, where we adjust the dimension along which we quantize the Key activations to better match the distribution; (ii) Pre-RoPE Key Quantization, where we quantize Key activations before the rotary positional embedding to mitigate its impact on quantization; (iii) Non-Uniform KV Cache Quantization, where we derive per-layer sensitivity-weighted non-uniform datatypes that better represent the distributions; and (iv) Per-Vector Dense-and-Sparse Quantization, where we isolate outliers separately for each vector to minimize skews in quantization ranges. By applying our method to the LLaMA, Llama-2, Llama-3, and Mistral models, we achieve < 0.1 perplexity degradation with 3-bit quantization on both Wikitext-2 and C4, outperforming existing approaches. Our method enables serving LLaMA-7B with a context length of up to 1 million on a single A100-80GB GPU and up to 10 million on an 8-GPU system. We develop custom CUDA kernels for KVQuant, showing that we can achieve up to ~1.7x speedups, compared to baseline fp16 matrix-vector multiplications, for the LLaMA-7B model.
Quantization is a crucial technique for deploying deep learning models on resource-constrained devices, such as embedded FPGAs. Prior efforts mostly focus on quantizing matrix multiplications, leaving other layers like BatchNorm or shortcuts in floating-point form, even though fixed-point arithmetic is more efficient on FPGAs. A common practice is to fine-tune a pre-trained model to fixed-point for FPGA deployment, but potentially degrading accuracy. This work presents QFX, a novel trainable fixed-point quantization approach that automatically learns the binary-point position during model training. Additionally, we introduce a multiplier-free quantization strategy within QFX to minimize DSP usage. QFX is implemented as a PyTorch-based library that efficiently emulates fixed-point arithmetic, supported by FPGA HLS, in a differentiable manner during backpropagation. With minimal effort, models trained with QFX can readily be deployed through HLS, producing the same numerical results as their software counterparts. Our evaluation shows that compared to post-training quantization, QFX can quantize models trained with element-wise layers quantized to fewer bits and achieve higher accuracy on both CIFAR-10 and ImageNet datasets. We further demonstrate the efficacy of multiplier-free quantization using a state-of-the-art binarized neural network accelerator designed for an embedded FPGA (AMD Xilinx Ultra96 v2). We plan to release QFX in open-source format.
Within-basket recommendation (WBR) refers to the task of recommending items to the end of completing a non-empty shopping basket during a shopping session. While the latest innovations in this space demonstrate remarkable performance improvement on benchmark datasets, they often overlook the complexity of user behaviors in practice, such as 1) co-existence of multiple shopping intentions, 2) multi-granularity of such intentions, and 3) interleaving behavior (switching intentions) in a shopping session. This paper presents Neural Pattern Associator (NPA), a deep item-association-mining model that explicitly models the aforementioned factors. Specifically, inspired by vector quantization, the NPA model learns to encode common user intentions (or item-combination patterns) as quantized representations (a.k.a. codebook), which permits identification of users's shopping intentions via attention-driven lookup during the reasoning phase. This yields coherent and self-interpretable recommendations. We evaluated the proposed NPA model across multiple extensive datasets, encompassing the domains of grocery e-commerce (shopping basket completion) and music (playlist extension), where our quantitative evaluations show that the NPA model significantly outperforms a wide range of existing WBR solutions, reflecting the benefit of explicitly modeling complex user intentions.
Weight quantization is an effective technique to compress deep neural networks for their deployment on edge devices with limited resources. Traditional loss-aware quantization methods commonly use the quantized gradient to replace the full-precision gradient. However, we discover that the gradient error will lead to an unexpected zig-zagging-like issue in the gradient descent learning procedures, where the gradient directions rapidly oscillate or zig-zag, and such issue seriously slows down the model convergence. Accordingly, this paper proposes a one-step forward and backtrack way for loss-aware quantization to get more accurate and stable gradient direction to defy this issue. During the gradient descent learning, a one-step forward search is designed to find the trial gradient of the next-step, which is adopted to adjust the gradient of current step towards the direction of fast convergence. After that, we backtrack the current step to update the full-precision and quantized weights through the current-step gradient and the trial gradient. A series of theoretical analysis and experiments on benchmark deep models have demonstrated the effectiveness and competitiveness of the proposed method, and our method especially outperforms others on the convergence performance.
This paper examines the quantization methods used in large-scale data analysis models and their hyperparameter choices. The recent surge in data analysis scale has significantly increased computational resource requirements. To address this, quantizing model weights has become a prevalent practice in data analysis applications such as deep learning. Quantization is particularly vital for deploying large models on devices with limited computational resources. However, the selection of quantization hyperparameters, like the number of bits and value range for weight quantization, remains an underexplored area. In this study, we employ the typical case analysis from statistical physics, specifically the replica method, to explore the impact of hyperparameters on the quantization of simple learning models. Our analysis yields three key findings: (i) an unstable hyperparameter phase, known as replica symmetry breaking, occurs with a small number of bits and a large quantization width; (ii) there is an optimal quantization width that minimizes error; and (iii) quantization delays the onset of overparameterization, helping to mitigate overfitting as indicated by the double descent phenomenon. We also discover that non-uniform quantization can enhance stability. Additionally, we develop an approximate message-passing algorithm to validate our theoretical results.
The remote wireless control of industrial systems is one of the major use cases for 5G and beyond systems: in these cases, the massive amounts of sensory information that need to be shared over the wireless medium may overload even high-capacity connections. Consequently, solving the effective communication problem by optimizing the transmission strategy to discard irrelevant information can provide a significant advantage, but is often a very complex task. In this work, we consider a prototypal system in which an observer must communicate its sensory data to a robot controlling a task (e.g., a mobile robot in a factory). We then model it as a remote Partially Observable Markov Decision Process (POMDP), considering the effect of adopting semantic and effective communication-oriented solutions on the overall system performance. We split the communication problem by considering an ensemble Vector Quantized Variational Autoencoder (VQ-VAE) encoding, and train a Deep Reinforcement Learning (DRL) agent to dynamically adapt the quantization level, considering both the current state of the environment and the memory of past messages. We tested the proposed approach on the well-known CartPole reference control problem, obtaining a significant performance increase over traditional approaches.
Large Language Models (LLMs) are difficult to fully fine-tune (e.g., with instructions or human feedback) due to their sheer number of parameters. A family of parameter-efficient sparse fine-tuning methods have proven promising in terms of performance but their memory requirements increase proportionally to the size of the LLMs. In this work, we scale sparse fine-tuning to state-of-the-art LLMs like LLaMA 2 7B and 13B. We propose SpIEL, a novel sparse fine-tuning method which, for a desired density level, maintains an array of parameter indices and the deltas of these parameters relative to their pretrained values. It iterates over: (a) updating the active deltas, (b) pruning indices (based on the change of magnitude of their deltas) and (c) regrowth of indices. For regrowth, we explore two criteria based on either the accumulated gradients of a few candidate parameters or their approximate momenta estimated using the efficient SM3 optimizer. We experiment with instruction-tuning of LLMs on standard dataset mixtures, finding that SpIEL is often superior to popular parameter-efficient fine-tuning methods like LoRA (low-rank adaptation) in terms of performance and comparable in terms of run time. We additionally show that SpIEL is compatible with both quantization and efficient optimizers, to facilitate scaling to ever-larger model sizes. We release the code for SpIEL at https://github.com/AlanAnsell/peft and for the instruction-tuning experiments at https://github.com/ducdauge/sft-llm.
Due to highly constrained computing power and memory, deploying 3D lidar-based detectors on edge devices equipped in autonomous vehicles and robots poses a crucial challenge. Being a convenient and straightforward model compression approach, Post-Training Quantization (PTQ) has been widely adopted in 2D vision tasks. However, applying it directly to 3D lidar-based tasks inevitably leads to performance degradation. As a remedy, we propose an effective PTQ method called LiDAR-PTQ, which is particularly curated for 3D lidar detection (both SPConv-based and SPConv-free). Our LiDAR-PTQ features three main components, \textbf{(1)} a sparsity-based calibration method to determine the initialization of quantization parameters, \textbf{(2)} a Task-guided Global Positive Loss (TGPL) to reduce the disparity between the final predictions before and after quantization, \textbf{(3)} an adaptive rounding-to-nearest operation to minimize the layerwise reconstruction error. Extensive experiments demonstrate that our LiDAR-PTQ can achieve state-of-the-art quantization performance when applied to CenterPoint (both Pillar-based and Voxel-based). To our knowledge, for the very first time in lidar-based 3D detection tasks, the PTQ INT8 model's accuracy is almost the same as the FP32 model while enjoying $3\times$ inference speedup. Moreover, our LiDAR-PTQ is cost-effective being $30\times$ faster than the quantization-aware training method. Code will be released at \url{https://github.com/StiphyJay/LiDAR-PTQ}.
Unsupervised image retrieval aims to learn the important visual characteristics without any given level to retrieve the similar images for a given query image. The Convolutional Neural Network (CNN)-based approaches have been extensively exploited with self-supervised contrastive learning for image hashing. However, the existing approaches suffer due to lack of effective utilization of global features by CNNs and biased-ness created by false negative pairs in the contrastive learning. In this paper, we propose a TransClippedCLR model by encoding the global context of an image using Transformer having local context through patch based processing, by generating the hash codes through product quantization and by avoiding the potential false negative pairs through clipped contrastive learning. The proposed model is tested with superior performance for unsupervised image retrieval on benchmark datasets, including CIFAR10, NUS-Wide and Flickr25K, as compared to the recent state-of-the-art deep models. The results using the proposed clipped contrastive learning are greatly improved on all datasets as compared to same backbone network with vanilla contrastive learning.
Secure two-party computation with homomorphic encryption (HE) protects data privacy with a formal security guarantee but suffers from high communication overhead. While previous works, e.g., Cheetah, Iron, etc, have proposed efficient HE-based protocols for different neural network (NN) operations, they still assume high precision, e.g., fixed point 37 bit, for the NN operations and ignore NNs' native robustness against quantization error. In this paper, we propose HEQuant, which features low-precision-quantization-aware optimization for the HE-based protocols. We observe the benefit of a naive combination of quantization and HE quickly saturates as bit precision goes down. Hence, to further improve communication efficiency, we propose a series of optimizations, including an intra-coefficient packing algorithm and a quantization-aware tiling algorithm, to simultaneously reduce the number and precision of the transferred data. Compared with prior-art HE-based protocols, e.g., CrypTFlow2, Cheetah, Iron, etc, HEQuant achieves $3.5\sim 23.4\times$ communication reduction and $3.0\sim 9.3\times$ latency reduction. Meanwhile, when compared with prior-art network optimization frameworks, e.g., SENet, SNL, etc, HEQuant also achieves $3.1\sim 3.6\times$ communication reduction.
How can we compress language models without sacrificing accuracy? The number of compression algorithms for language models is rapidly growing to benefit from remarkable advances of recent language models without side effects due to the gigantic size of language models, such as increased carbon emissions and expensive maintenance fees. While numerous compression algorithms have shown remarkable progress in compressing language models, it ironically becomes challenging to capture emerging trends and identify the fundamental concepts underlying them due to the excessive number of algorithms. In this paper, we survey and summarize diverse compression algorithms including pruning, quantization, knowledge distillation, low-rank approximation, parameter sharing, and efficient architecture design. We not only summarize the overall trend of diverse compression algorithms but also select representative algorithms and provide in-depth analyses of them. We discuss the value of each category of compression algorithms, and the desired properties of low-cost compression algorithms which have a significant impact due to the emergence of large language models. Finally, we introduce promising future research topics based on our survey results.
Vector quantization is a fundamental operation for data compression and vector search. To obtain high accuracy, multi-codebook methods represent each vector using codewords across several codebooks. Residual quantization (RQ) is one such method, which iteratively quantizes the error of the previous step. While the error distribution is dependent on previously-selected codewords, this dependency is not accounted for in conventional RQ as it uses a fixed codebook per quantization step. In this paper, we propose QINCo, a neural RQ variant that constructs specialized codebooks per step that depend on the approximation of the vector from previous steps. Experiments show that QINCo outperforms state-of-the-art methods by a large margin on several datasets and code sizes. For example, QINCo achieves better nearest-neighbor search accuracy using 12-byte codes than the state-of-the-art UNQ using 16 bytes on the BigANN1M and Deep1M datasets.
While vision transformers (ViTs) have shown great potential in computer vision tasks, their intense computation and memory requirements pose challenges for practical applications. Existing post-training quantization methods leverage value redistribution or specialized quantizers to address the non-normal distribution in ViTs. However, without considering the asymmetry in activations and relying on hand-crafted settings, these methods often struggle to maintain performance under low-bit quantization. To overcome these challenges, we introduce SmoothQuant with bias term (SQ-b) to alleviate the asymmetry issue and reduce the clamping loss. We also introduce optimal scaling factor ratio search (OPT-m) to determine quantization parameters by a data-dependent mechanism automatically. To further enhance the compressibility, we incorporate the above-mentioned techniques and propose a mixed-precision post-training quantization framework for vision transformers (MPTQ-ViT). We develop greedy mixed-precision quantization (Greedy MP) to allocate layer-wise bit-width considering both model performance and compressibility. Our experiments on ViT, DeiT, and Swin demonstrate significant accuracy improvements compared with SOTA on the ImageNet dataset. Specifically, our proposed methods achieve accuracy improvements ranging from 0.90% to 23.35% on 4-bit ViTs with single-precision and from 3.82% to 78.14% on 5-bit fully quantized ViTs with mixed-precision.