Detection and description of keypoints in images is a fundamental component of a wide range of tasks such as Simultaneous Localization And Mapping (SLAM), image alignment and structure from motion (SfM). Efficient computation of these features is crucial for real-time applications and has been addressed by multiple handcrafted algorithms and, recently, by deep neural network-based detectors. Learned detectors achieve high detection performance, but pose high computational requirements, making them slow and impractical for low-power resource constraint platforms. This paper presents a quantized neural keypoint detector and descriptor optimized for edge devices exploiting two recent AI platforms such as MAX78000 by Analog Devices and the Coral AI USB accelerator from Google. To accommodate the diverse constraints and requirements of various applications, we propose and evaluate two model architectures (KP2DtinySmall and KP2DtinyFast) and deploy them on the aforementioned platforms using full 8-bit integer quantization. Furthermore, we extensively evaluate these models in terms of power, latency and accuracy, reporting results on three image sizes (88x88, 320x240 and 640x480), evaluating both quantized and non-quantized models. Fully quantized, KP2DtinySmall reduces network size by a factor of 54x while improving homographic estimation accuracy on 88x88 images on the most stringent threshold (Correctness d1) by 32.4% (0.550) and on 320x240 images by 10.7% (0.648) compared to the KeypointNet architecture by Yang You et. al. This result is achieved by designing a new network with low power platforms in mind, particularly addressing the lower resolution by increasing the density of detectable features. Deployed on the MAX78000 MCU, inference of low-resolution images is run at 59 FPS, consuming 1.1 mJ per image. On the Coral usb accelerator, KP2DtinyFast runs inference on low-resolution images at 527 FPS consuming 3.1 mJ, on high resolution it achieves 70 FPS at 19.9 mJ per inference.
Deployment of Large Language Models (LLMs) has major computational costs, due to their rapidly expanding size. Compression of LLMs reduces the memory footprint, latency, and energy required for their inference. Post-training Quantization (PTQ) techniques have been developed to compress LLMs while avoiding expensive re-training. Most PTQ approaches formulate the quantization error based on a layer-wise Euclidean loss, ignoring the model output. Then, each layer is calibrated using its layer-wise Hessian to update the weights towards minimizing the quantization error. The Hessian is also used for detecting the most salient weights to quantization. Such PTQ approaches are prone to accuracy drop in low-precision quantization. We propose Output-adaptive Calibration (OAC) to incorporate the model output in the calibration process. We formulate the quantization error based on the distortion of the output cross-entropy loss. OAC approximates the output-adaptive Hessian for each layer under reasonable assumptions to reduce the computational complexity. The output-adaptive Hessians are used to update the weight matrices and detect the salient weights towards maintaining the model output. Our proposed method outperforms the state-of-the-art baselines such as SpQR and BiLLM, especially, at extreme low-precision (2-bit and binary) quantization.
Lipschitz bounds on neural network parameterizations are important to establish generalization, quantization or pruning guarantees, as they control the robustness of the network with respect to parameter changes. Yet, there are few Lipschitz bounds with respect to parameters in the literature, and existing ones only apply to simple feedforward architectures, while also failing to capture the intrinsic rescaling-symmetries of ReLU networks. This paper proves a new Lipschitz bound in terms of the so-called path-metrics of the parameters. Since this bound is intrinsically invariant with respect to the rescaling symmetries of the networks, it sharpens previously known Lipschitz bounds. It is also, to the best of our knowledge, the first bound of its kind that is broadly applicable to modern networks such as ResNets, VGGs, U-nets, and many more.
Large language models (LLMs) achieve remarkable performance in natural language understanding but require substantial computation and memory resources. Post-training quantization (PTQ) is a powerful compression technique extensively investigated in LLMs. However, existing PTQ methods are still not ideal in terms of accuracy and efficiency, especially with below 4 bit-widths. Standard PTQ methods using group-wise quantization suffer difficulties in quantizing LLMs accurately to such low-bit, but advanced methods remaining high-precision weights element-wisely are hard to realize their theoretical hardware efficiency. This paper presents a Salience-Driven Mixed-Precision Quantization scheme for LLMs, namely SliM-LLM. The scheme exploits the salience distribution of weights to determine optimal bit-width and quantizers for accurate LLM quantization, while aligning bit-width partition to groups for compact memory usage and fast integer inference. Specifically, the proposed SliM-LLM mainly relies on two novel techniques: (1) Salience-Determined Bit Allocation utilizes the clustering characteristics of salience distribution to allocate the bit-widths of each group, increasing the accuracy of quantized LLMs and maintaining the inference efficiency; (2) Salience-Weighted Quantizer Calibration optimizes the parameters of the quantizer by considering the element-wise salience within the group, balancing the maintenance of salient information and minimization of errors. Comprehensive experiments show that SliM-LLM significantly improves the accuracy of LLMs at ultra-low bits, e.g., 2-bit LLaMA-7B achieves a 5.5-times memory-saving than original model on NVIDIA A800 GPUs, and 48% decrease of perplexity compared to the state-of-the-art gradient-free PTQ method. Moreover, SliM-LLM+, which is integrated from the extension of SliM-LLM with gradient-based quantizers, further reduces perplexity by 35.1%.
As the size and context length of Large Language Models (LLMs) grow, weight-activation quantization has emerged as a crucial technique for efficient deployment of LLMs. Compared to weight-only quantization, weight-activation quantization presents greater challenges due to the presence of outliers in activations. Existing methods have made significant progress by exploring mixed-precision quantization and outlier suppression. However, these methods primarily focus on optimizing the results of single matrix multiplication, neglecting the bidirectional propagation of quantization errors in LLMs. Specifically, errors accumulate vertically within the same token through layers, and diffuse horizontally across different tokens due to self-attention mechanisms. To address this issue, we introduce BiSup, a Bidirectional quantization error Suppression method. By constructing appropriate optimizable parameter spaces, BiSup utilizes a small amount of data for quantization-aware parameter-efficient fine-tuning to suppress the error vertical accumulation. Besides, BiSup employs prompt mixed-precision quantization strategy, which preserves high precision for the key-value cache of system prompts, to mitigate the error horizontal diffusion. Extensive experiments on Llama and Qwen families demonstrate that BiSup can improve performance over two state-of-the-art methods (the average WikiText2 perplexity decreases from 13.26 to 9.41 for Atom and from 14.33 to 7.85 for QuaRot under the W3A3-g128 configuration), further facilitating the practical applications of low-bit weight-activation quantization.
Existing deep learning methods have made significant progress in gait recognition. Typically, appearance-based models binarize inputs into silhouette sequences. However, mainstream quantization methods prioritize minimizing task loss over quantization error, which is detrimental to gait recognition with binarized inputs. Minor variations in silhouette sequences can be diminished in the network's intermediate layers due to the accumulation of quantization errors. To address this, we propose a differentiable soft quantizer, which better simulates the gradient of the round function during backpropagation. This enables the network to learn from subtle input perturbations. However, our theoretical analysis and empirical studies reveal that directly applying the soft quantizer can hinder network convergence. We further refine the training strategy to ensure convergence while simulating quantization errors. Additionally, we visualize the distribution of outputs from different samples in the feature space and observe significant changes compared to the full precision network, which harms performance. Based on this, we propose an Inter-class Distance-guided Distillation (IDD) strategy to preserve the relative distance between the embeddings of samples with different labels. Extensive experiments validate the effectiveness of our approach, demonstrating state-of-the-art accuracy across various settings and datasets. The code will be made publicly available.
Semantic communication has undergone considerable evolution due to the recent rapid development of artificial intelligence (AI), significantly enhancing both communication robustness and efficiency. Despite these advancements, most current semantic communication methods for image transmission pay little attention to the differing importance of objects and backgrounds in images. To address this issue, we propose a novel scheme named ASCViT-JSCC, which utilizes vision transformers (ViTs) integrated with an orthogonal frequency division multiplexing (OFDM) system. This scheme adaptively allocates bandwidth for objects and backgrounds in images according to the importance order of different parts determined by object detection of you only look once version 5 (YOLOv5) and feature points detection of scale invariant feature transform (SIFT). Furthermore, the proposed scheme adheres to digital modulation standards by incorporating quantization modules. We validate this approach through an over-the-air (OTA) testbed named intelligent communication prototype validation platform (ICP) based on a software-defined radio (SDR) and NVIDIA embedded kits. Our findings from both simulations and practical measurements show that ASCViT-JSCC significantly preserves objects in images and enhances reconstruction quality compared to existing methods.
We introduce ReALLM, a novel approach for compression and memory-efficient adaptation of pre-trained language models that encompasses most of the post-training quantization and fine-tuning methods for a budget of <4 bits. Pre-trained matrices are decomposed into a high-precision low-rank component and a vector-quantized latent representation (using an autoencoder). During the fine-tuning step, only the low-rank components are updated. Our results show that pre-trained matrices exhibit different patterns. ReALLM adapts the shape of the encoder (small/large embedding, high/low bit VQ, etc.) to each matrix. ReALLM proposes to represent each matrix with a small embedding on $b$ bits and a neural decoder model $\mathcal{D}_\phi$ with its weights on $b_\phi$ bits. The decompression of a matrix requires only one embedding and a single forward pass with the decoder. Our weight-only quantization algorithm yields the best results on language generation tasks (C4 and WikiText-2) for a budget of $3$ bits without any training. With a budget of $2$ bits, ReALLM achieves state-of-the art performance after fine-tuning on a small calibration dataset.
Generative retrieval, a promising new paradigm in information retrieval, employs a seq2seq model to encode document features into parameters and decode relevant document identifiers (IDs) based on search queries. Existing generative retrieval solutions typically rely on a preprocessing stage to pre-define document IDs, which can suffer from a semantic gap between these IDs and the retrieval task. However, end-to-end training for both ID assignments and retrieval tasks is challenging due to the long-tailed distribution characteristics of real-world data, resulting in inefficient and unbalanced ID space utilization. To address these issues, we propose ASI++, a novel fully end-to-end generative retrieval method that aims to simultaneously learn balanced ID assignments and improve retrieval performance. ASI++ builds on the fully end-to-end training framework of vanilla ASI and introduces several key innovations. First, a distributionally balanced criterion addresses the imbalance in ID assignments, promoting more efficient utilization of the ID space. Next, a representation bottleneck criterion enhances dense representations to alleviate bottlenecks in learning ID assignments. Finally, an information consistency criterion integrates these processes into a joint optimization framework grounded in information theory. We further explore various module structures for learning ID assignments, including neural quantization, differentiable product quantization, and residual quantization. Extensive experiments on both public and industrial datasets demonstrate the effectiveness of ASI++ in improving retrieval performance and achieving balanced ID assignments.
This paper presents a novel approach to enhance communication efficiency in federated learning through clipped uniform quantization. By leveraging optimal clipping thresholds and client-specific adaptive quantization schemes, the proposed method significantly reduces bandwidth and memory requirements for model weight transmission between clients and the server while maintaining competitive accuracy. We investigate the effects of symmetric clipping and uniform quantization on model performance, emphasizing the role of stochastic quantization in mitigating artifacts and improving robustness. Extensive simulations demonstrate that the method achieves near-full-precision performance with substantial communication savings. Moreover, the proposed approach facilitates efficient weight averaging based on the inverse of the mean squared quantization errors, effectively balancing the trade-off between communication efficiency and model accuracy. Moreover, in contrast to federated averaging, this design obviates the need to disclose client-specific data volumes to the server, thereby enhancing client privacy. Comparative analysis with conventional quantization methods further confirms the efficacy of the proposed scheme.
The ever-growing computational complexity of Large Language Models (LLMs) necessitates efficient deployment strategies. The current state-of-the-art approaches for Post-training Quantization (PTQ) often require calibration to achieve the desired accuracy. This paper presents AdpQ, a novel zero-shot adaptive PTQ method for LLMs that achieves the state-of-the-art performance in low-precision quantization (e.g. 3-bit) without requiring any calibration data. Inspired by Adaptive LASSO regression model, our proposed approach tackles the challenge of outlier activations by separating salient weights using an adaptive soft-thresholding method. Guided by Adaptive LASSO, this method ensures that the quantized weights distribution closely follows the originally trained weights and eliminates the need for calibration data entirely, setting our method apart from popular approaches such as SpQR and AWQ. Furthermore, our method offers an additional benefit in terms of privacy preservation by eliminating any calibration or training data. We also delve deeper into the information-theoretic underpinnings of the proposed method. We demonstrate that it leverages the Adaptive LASSO to minimize the Kullback-Leibler divergence between the quantized weights and the originally trained weights. This minimization ensures the quantized model retains the Shannon information content of the original model to a great extent, guaranteeing efficient deployment without sacrificing accuracy or information. Our results achieve the same accuracy as the existing methods on various LLM benchmarks while the quantization time is reduced by at least 10x, solidifying our contribution to efficient and privacy-preserving LLM deployment.
Central to the Transformer architectures' effectiveness is the self-attention mechanism, a function that maps queries, keys, and values into a high-dimensional vector space. However, training the attention weights of queries, keys, and values is non-trivial from a state of random initialization. In this paper, we propose two methods. (i) We first address the initialization problem of Vision Transformers by introducing a simple, yet highly innovative, initialization approach utilizing discrete cosine transform (DCT) coefficients. Our proposed DCT-based \textit{attention} initialization marks a significant gain compared to traditional initialization strategies; offering a robust foundation for the attention mechanism. Our experiments reveal that the DCT-based initialization enhances the accuracy of Vision Transformers in classification tasks. (ii) We also recognize that since DCT effectively decorrelates image information in the frequency domain, this decorrelation is useful for compression because it allows the quantization step to discard many of the higher-frequency components. Based on this observation, we propose a novel DCT-based compression technique for the attention function of Vision Transformers. Since high-frequency DCT coefficients usually correspond to noise, we truncate the high-frequency DCT components of the input patches. Our DCT-based compression reduces the size of weight matrices for queries, keys, and values. While maintaining the same level of accuracy, our DCT compressed Swin Transformers obtain a considerable decrease in the computational overhead.
There has been significant interest in "extreme" compression of large language models (LLMs), i.e., to 1-2 bits per parameter, which allows such models to be executed efficiently on resource-constrained devices. Existing work focused on improved one-shot quantization techniques and weight representations; yet, purely post-training approaches are reaching diminishing returns in terms of the accuracy-vs-bit-width trade-off. State-of-the-art quantization methods such as QuIP# and AQLM include fine-tuning (part of) the compressed parameters over a limited amount of calibration data; however, such fine-tuning techniques over compressed weights often make exclusive use of straight-through estimators (STE), whose performance is not well-understood in this setting. In this work, we question the use of STE for extreme LLM compression, showing that it can be sub-optimal, and perform a systematic study of quantization-aware fine-tuning strategies for LLMs. We propose PV-Tuning - a representation-agnostic framework that generalizes and improves upon existing fine-tuning strategies, and provides convergence guarantees in restricted cases. On the practical side, when used for 1-2 bit vector quantization, PV-Tuning outperforms prior techniques for highly-performant models such as Llama and Mistral. Using PV-Tuning, we achieve the first Pareto-optimal quantization for Llama 2 family models at 2 bits per parameter.
Modern large language models (LLMs) have established state-of-the-art performance through architectural improvements, but still require significant computational cost for inference. In an effort to reduce the inference cost, post-training quantization (PTQ) has become a popular approach, quantizing weights and activations to lower precision, such as INT8. In this paper, we reveal the challenges of activation quantization in GLU variants, which are widely used in feed-forward network (FFN) of modern LLMs, such as LLaMA family. The problem is that severe local quantization errors, caused by excessive magnitudes of activation in GLU variants, significantly degrade the performance of the quantized LLM. We denote these activations as activation spikes. Our further observations provide a systematic pattern of activation spikes: 1) The activation spikes occur in the FFN of specific layers, particularly in the early and late layers, 2) The activation spikes are dedicated to a couple of tokens, rather than being shared across a sequence. Based on our observations, we propose two empirical methods, Quantization-free Module (QFeM) and Quantization-free Prefix (QFeP), to isolate the activation spikes during quantization. Our extensive experiments validate the effectiveness of the proposed methods for the activation quantization, especially with coarse-grained scheme, of latest LLMs with GLU variants, including LLaMA-2/3, Mistral, Mixtral, SOLAR, and Gemma. In particular, our methods enhance the current alleviation techniques (e.g., SmoothQuant) that fail to control the activation spikes. Code is available at https://github.com/onnoo/activation-spikes.
Real world re-identfication (ReID) algorithms aim to map new observations of an object to previously recorded instances. These systems are often constrained by quantity and size of the stored embeddings. To combat this scaling problem, we attempt to shrink the size of these vectors by using a variety of compression techniques. In this paper, we benchmark quantization-aware-training along with three different dimension reduction methods: iterative structured pruning, slicing the embeddings at initialize, and using low rank embeddings. We find that ReID embeddings can be compressed by up to 96x with minimal drop in performance. This implies that modern re-identification paradigms do not fully leverage the high dimensional latent space, opening up further research to increase the capabilities of these systems.
``Learning to hash'' is a practical solution for efficient retrieval, offering fast search speed and low storage cost. It is widely applied in various applications, such as image-text cross-modal search. In this paper, we explore the potential of enhancing the performance of learning to hash with the proliferation of powerful large pre-trained models, such as Vision-Language Pre-training (VLP) models. We introduce a novel method named Distillation for Cross-Modal Quantization (DCMQ), which leverages the rich semantic knowledge of VLP models to improve hash representation learning. Specifically, we use the VLP as a `teacher' to distill knowledge into a `student' hashing model equipped with codebooks. This process involves the replacement of supervised labels, which are composed of multi-hot vectors and lack semantics, with the rich semantics of VLP. In the end, we apply a transformation termed Normalization with Paired Consistency (NPC) to achieve a discriminative target for distillation. Further, we introduce a new quantization method, Product Quantization with Gumbel (PQG) that promotes balanced codebook learning, thereby improving the retrieval performance. Extensive benchmark testing demonstrates that DCMQ consistently outperforms existing supervised cross-modal hashing approaches, showcasing its significant potential.
We demonstrate generating HDR images using the concerted action of multiple black-box, pre-trained LDR image diffusion models. Relying on a pre-trained LDR generative diffusion models is vital as, first, there is no sufficiently large HDR image dataset available to re-train them, and, second, even if it was, re-training such models is impossible for most compute budgets. Instead, we seek inspiration from the HDR image capture literature that traditionally fuses sets of LDR images, called "exposure brackets'', to produce a single HDR image. We operate multiple denoising processes to generate multiple LDR brackets that together form a valid HDR result. The key to making this work is to introduce a consistency term into the diffusion process to couple the brackets such that they agree across the exposure range they share while accounting for possible differences due to the quantization error. We demonstrate state-of-the-art unconditional and conditional or restoration-type (LDR2HDR) generative modeling results, yet in HDR.
Learning discrete representations with vector quantization (VQ) has emerged as a powerful approach in various generative models. However, most VQ-based models rely on a single, fixed-rate codebook, requiring extensive retraining for new bitrates or efficiency requirements. We introduce Rate-Adaptive Quantization (RAQ), a multi-rate codebook adaptation framework for VQ-based generative models. RAQ applies a data-driven approach to generate variable-rate codebooks from a single baseline VQ model, enabling flexible tradeoffs between compression and reconstruction fidelity. Additionally, we provide a simple clustering-based procedure for pre-trained VQ models, offering an alternative when retraining is infeasible. Our experiments show that RAQ performs effectively across multiple rates, often outperforming conventional fixed-rate VQ baselines. By enabling a single system to seamlessly handle diverse bitrate requirements, RAQ extends the adaptability of VQ-based generative models and broadens their applicability to data compression, reconstruction, and generation tasks.
Vector quantization (VQ) is a key technique in high-resolution and high-fidelity image synthesis, which aims to learn a codebook to encode an image with a sequence of discrete codes and then generate an image in an auto-regression manner. Although existing methods have shown superior performance, most methods prefer to learn a single-modal codebook (\emph{e.g.}, image), resulting in suboptimal performance when the codebook is applied to multi-modal downstream tasks (\emph{e.g.}, text-to-image, image captioning) due to the existence of modal gaps. In this paper, we propose a novel language-guided codebook learning framework, called LG-VQ, which aims to learn a codebook that can be aligned with the text to improve the performance of multi-modal downstream tasks. Specifically, we first introduce pre-trained text semantics as prior knowledge, then design two novel alignment modules (\emph{i.e.}, Semantic Alignment Module, and Relationship Alignment Module) to transfer such prior knowledge into codes for achieving codebook text alignment. In particular, our LG-VQ method is model-agnostic, which can be easily integrated into existing VQ models. Experimental results show that our method achieves superior performance on reconstruction and various multi-modal downstream tasks.
eXmY is a novel data type for quantization of ML models. It supports both arbitrary bit widths and arbitrary integer and floating point formats. For example, it seamlessly supports 3, 5, 6, 7, 9 bit formats. For a specific bit width, say 7, it defines all possible formats e.g. e0m6, e1m5, e2m4, e3m3, e4m2, e5m1 and e6m0. For non-power of two bit widths e.g. 5, 6, 7, we created a novel encoding and decoding scheme which achieves perfect compression, byte addressability and is amenable to sharding and vector processing. We implemented libraries for emulation, encoding and decoding tensors and checkpoints in C++, TensorFlow, JAX and PAX. For optimal performance, the codecs use SIMD instructions on CPUs and vector instructions on TPUs and GPUs. eXmY is also a technique and exploits the statistical distribution of exponents in tensors. It can be used to quantize weights, static and dynamic activations, gradients, master weights and optimizer state. It can reduce memory (CPU DRAM and accelerator HBM), network and disk storage and transfers. It can increase multi tenancy and accelerate compute. eXmY has been deployed in production for almost 2 years.