Human pose estimation on medium and small scales has long been a significant challenge in this field. Most existing methods focus on restoring high-resolution feature maps by stacking multiple costly deconvolutional layers or by continuously aggregating semantic information from low-resolution feature maps while maintaining high-resolution ones, which can lead to information redundancy. Additionally, due to quantization errors, heatmap-based methods have certain disadvantages in accurately locating keypoints of medium and small-scale human figures. In this paper, we propose HRPVT, which utilizes PVT v2 as the backbone to model long-range dependencies. Building on this, we introduce the High-Resolution Pyramid Module (HRPM), designed to generate higher quality high-resolution representations by incorporating the intrinsic inductive biases of Convolutional Neural Networks (CNNs) into the high-resolution feature maps. The integration of HRPM enhances the performance of pure transformer-based models for human pose estimation at medium and small scales. Furthermore, we replace the heatmap-based method with SimCC approach, which eliminates the need for costly upsampling layers, thereby allowing us to allocate more computational resources to HRPM. To accommodate models with varying parameter scales, we have developed two insertion strategies of HRPM, each designed to enhancing the model's ability to perceive medium and small-scale human poses from two distinct perspectives.
Fine-tuning pre-trained diffusion models under limited budgets has gained great success. In particular, the recent advances that directly fine-tune the quantized weights using Low-rank Adaptation (LoRA) further reduces training costs. Despite these progress, we point out that existing adaptation recipes are not inference-efficient. Specifically, additional post-training quantization (PTQ) on tuned weights is needed during deployment, which results in noticeable performance drop when the bit-width is low. Based on this observation, we introduce IntLoRA, which adapts quantized diffusion models with integer-type low-rank parameters, to include inference efficiency during tuning. Specifically, IntLoRA enables pre-trained weights to remain quantized during training, facilitating fine-tuning on consumer-level GPUs. During inference, IntLoRA weights can be seamlessly merged into pre-trained weights to directly obtain quantized downstream weights without PTQ. Extensive experiments show our IntLoRA achieves significant speedup on both training and inference without losing performance.
Last few years have seen unprecedented advances in capabilities of Large Language Models (LLMs). These advancements promise to deeply benefit a vast array of application domains. However, due to their immense size, performing inference with LLMs is both costly and slow. Consequently, a plethora of recent work has proposed strategies to enhance inference efficiency, e.g., quantization, pruning, and caching. These acceleration strategies reduce the inference cost and latency, often by several factors, while maintaining much of the predictive performance measured via common benchmarks. In this work, we explore another critical aspect of LLM performance: demographic bias in model generations due to inference acceleration optimizations. Using a wide range of metrics, we probe bias in model outputs from a number of angles. Analysis of outputs before and after inference acceleration shows significant change in bias. Worryingly, these bias effects are complex and unpredictable. A combination of an acceleration strategy and bias type may show little bias change in one model but may lead to a large effect in another. Our results highlight a need for in-depth and case-by-case evaluation of model bias after it has been modified to accelerate inference.
This paper introduces a robust stress detection system utilizing a Convolutional Neural Network (CNN) designed for the analysis of Photoplethysmogram (PPG) signals. Employing the WESAD dataset, we applied Continuous Wavelet Transform (CWT) to extract informative features from wrist PPG signals, demonstrating enhanced stress detection and learning compared to conventional techniques. Notably, the CNN achieved an impressive accuracy of 93.7% after five epochs, post-implementation on a resource-constrained microcontroller. The optimization process, including pruning and Post-Train Quantization, was crucial to reduce the model size to 1.6 megabytes, overcoming the microcontroller's limited resources of 2 megabytes of Flash memory and 512 kilobytes of RAM. This optimized model not only addresses resource constraints but also outperforms traditional signal processing methods, positioning it as a promising solution for real-time stress monitoring on wearable devices.
Distributed optimization finds many applications in machine learning, signal processing, and control systems. In these real-world applications, the constraints of communication networks, particularly limited bandwidth, necessitate implementing quantization techniques. In this paper, we propose distributed optimization dynamics over multi-agent networks subject to logarithmically quantized data transmission. Under this condition, data exchange benefits from representing smaller values with more bits and larger values with fewer bits. As compared to uniform quantization, this allows for higher precision in representing near-optimal values and more accuracy of the distributed optimization algorithm. The proposed optimization dynamics comprise a primary state variable converging to the optimizer and an auxiliary variable tracking the objective function's gradient. Our setting accommodates dynamic network topologies, resulting in a hybrid system requiring convergence analysis using matrix perturbation theory and eigenspectrum analysis.
Conventional digitization based on the Shannon-Nyquist method, implemented via analog-to-digital converters (ADCs), faces fundamental limitations. High-dynamic-range (HDR) signals often get clipped or saturated in practice. Given a fixed bit budget, one must choose between minimizing quantization noise or accommodating HDR inputs. The Unlimited Sensing Framework (USF) eliminates saturation by incorporating nonlinear folding in analog hardware, resulting in modulo signals. Quantizing or digitizing modulo signals enables low quantization noise as the modulo representation maps HDR signals into low-dynamic-range (LDR) samples. In the context of USF, the core innovation of this paper is a novel, low-cost, integrator-based efficient modulo ADC hardware implementation that imposes no restrictions on folding rates, enabling significantly HDR capture. The feasibility of this design is demonstrated by hardware experiments showcasing clear advantages across different quantitative performance metrics. These include capturing HDR signals with a 60-fold increase in dynamic range, achieving up to 5 Effective Number of Bits (ENOBs), and improving Signal-to-Noise and Distortion (SINAD) by 30 dB.
Small Language Models (SLMs) have become increasingly important due to their efficiency and performance to perform various language tasks with minimal computational resources, making them ideal for various settings including on-device, mobile, edge devices, among many others. In this article, we present a comprehensive survey on SLMs, focusing on their architectures, training techniques, and model compression techniques. We propose a novel taxonomy for categorizing the methods used to optimize SLMs, including model compression, pruning, and quantization techniques. We summarize the benchmark datasets that are useful for benchmarking SLMs along with the evaluation metrics commonly used. Additionally, we highlight key open challenges that remain to be addressed. Our survey aims to serve as a valuable resource for researchers and practitioners interested in developing and deploying small yet efficient language models.
Continual learning requires to overcome catastrophic forgetting when training a single model on a sequence of tasks. Recent top-performing approaches are prompt-based methods that utilize a set of learnable parameters (i.e., prompts) to encode task knowledge, from which appropriate ones are selected to guide the fixed pre-trained model in generating features tailored to a certain task. However, existing methods rely on predicting prompt identities for prompt selection, where the identity prediction process cannot be optimized with task loss. This limitation leads to sub-optimal prompt selection and inadequate adaptation of pre-trained features for a specific task. Previous efforts have tried to address this by directly generating prompts from input queries instead of selecting from a set of candidates. However, these prompts are continuous, which lack sufficient abstraction for task knowledge representation, making them less effective for continual learning. To address these challenges, we propose VQ-Prompt, a prompt-based continual learning method that incorporates Vector Quantization (VQ) into end-to-end training of a set of discrete prompts. In this way, VQ-Prompt can optimize the prompt selection process with task loss and meanwhile achieve effective abstraction of task knowledge for continual learning. Extensive experiments show that VQ-Prompt outperforms state-of-the-art continual learning methods across a variety of benchmarks under the challenging class-incremental setting. The code is available at \href{https://github.com/jiaolifengmi/VQ-Prompt}{this https URL}.
We study the impact of a standard practice in compressing foundation vision-language models - quantization - on the models' ability to produce socially-fair outputs. In contrast to prior findings with unimodal models that compression consistently amplifies social biases, our extensive evaluation of four quantization settings across three datasets and three CLIP variants yields a surprising result: while individual models demonstrate bias, we find no consistent change in bias magnitude or direction across a population of compressed models due to quantization.
Large-scale recommendation models are currently the dominant workload for many large Internet companies. These recommenders are characterized by massive embedding tables that are sparsely accessed by the index for user and item features. The size of these 1TB+ tables imposes a severe memory bottleneck for the training and inference of recommendation models. In this work, we propose a novel recommendation framework that is small, powerful, and efficient to run and train, based on the state-of-the-art Deep Learning Recommendation Model (DLRM). The proposed framework makes inference more efficient on the cloud servers, explores the possibility of deploying powerful recommenders on smaller edge devices, and optimizes the workload of the communication overhead in distributed training under the data parallelism settings. Specifically, we show that quantization-aware training (QAT) can impose a strong regularization effect to mitigate the severe overfitting issues suffered by DLRMs. Consequently, we achieved INT4 quantization of DLRM models without any accuracy drop. We further propose two techniques that improve and accelerate the conventional QAT workload specifically for the embedding tables in the recommendation models. Furthermore, to achieve efficient training, we quantize the gradients of the embedding tables into INT8 on top of the well-supported specified sparsification. We show that combining gradient sparsification and quantization together significantly reduces the amount of communication. Briefly, DQRM models with INT4 can achieve 79.07% accuracy on Kaggle with 0.27 GB model size, and 81.21% accuracy on the Terabyte dataset with 1.57 GB, which even outperform FP32 DLRMs that have much larger model sizes (2.16 GB on Kaggle and 12.58 on Terabyte).
The performance of neural networks improves when more parameters are used. However, the model sizes are constrained by the available on-device memory during training and inference. Although applying techniques like quantization can alleviate the constraint, they suffer from performance degradation. In this work, we introduce NeuZip, a new weight compression scheme based on the entropy of floating-point numbers in neural networks. With NeuZip, we are able to achieve memory-efficient training and inference without sacrificing performance. Notably, we significantly reduce the memory footprint of training a Llama-3 8B model from 31GB to less than 16GB, while keeping the training dynamics fully unchanged. In inference, our method can reduce memory usage by more than half while maintaining near-lossless performance. Our code is publicly available.
Generative adversarial networks (GANs) learn a latent space whose samples can be mapped to real-world images. Such latent spaces are difficult to interpret. Some earlier supervised methods aim to create an interpretable latent space or discover interpretable directions that require exploiting data labels or annotated synthesized samples for training. However, we propose using a modification of vector quantization called space-filling vector quantization (SFVQ), which quantizes the data on a piece-wise linear curve. SFVQ can capture the underlying morphological structure of the latent space and thus make it interpretable. We apply this technique to model the latent space of pretrained StyleGAN2 and BigGAN networks on various datasets. Our experiments show that the SFVQ curve yields a general interpretable model of the latent space that determines which part of the latent space corresponds to what specific generative factors. Furthermore, we demonstrate that each line of SFVQ's curve can potentially refer to an interpretable direction for applying intelligible image transformations. We also showed that the points located on an SFVQ line can be used for controllable data augmentation.
In this work, we re-formulate the model compression problem into the customized compensation problem: Given a compressed model, we aim to introduce residual low-rank paths to compensate for compression errors under customized requirements from users (e.g., tasks, compression ratios), resulting in greater flexibility in adjusting overall capacity without being constrained by specific compression formats. However, naively applying SVD to derive residual paths causes suboptimal utilization of the low-rank representation capacity. Instead, we propose Training-free Eigenspace Low-Rank Approximation (EoRA), a method that directly minimizes compression-induced errors without requiring gradient-based training, achieving fast optimization in minutes using a small amount of calibration data. EoRA projects compression errors into the eigenspace of input activations, leveraging eigenvalues to effectively prioritize the reconstruction of high-importance error components. Moreover, EoRA can be seamlessly integrated with fine-tuning and quantization to further improve effectiveness and efficiency. EoRA consistently outperforms previous methods in compensating errors for compressed LLaMA2/3 models on various tasks, such as language generation, commonsense reasoning, and math reasoning tasks (e.g., 31.31%/12.88% and 9.69% improvements on ARC-Easy/ARC-Challenge and MathQA when compensating LLaMA3-8B that is quantized to 4-bit and pruned to 2:4 sparsity). EoRA offers a scalable, training-free solution to compensate for compression errors, making it a powerful tool to deploy LLMs in various capacity and efficiency requirements.
Current multimodal recommendation models have extensively explored the effective utilization of multimodal information; however, their reliance on ID embeddings remains a performance bottleneck. Even with the assistance of multimodal information, optimizing ID embeddings remains challenging for ID-based Multimodal Recommender when interaction data is sparse. Furthermore, the unique nature of item-specific ID embeddings hinders the information exchange among related items and the spatial requirement of ID embeddings increases with the scale of item. Based on these limitations, we propose an ID-free MultimOdal TOken Representation scheme named MOTOR that represents each item using learnable multimodal tokens and connects them through shared tokens. Specifically, we first employ product quantization to discretize each item's multimodal features (e.g., images, text) into discrete token IDs. We then interpret the token embeddings corresponding to these token IDs as implicit item features, introducing a new Token Cross Network to capture the implicit interaction patterns among these tokens. The resulting representations can replace the original ID embeddings and transform the original ID-based multimodal recommender into ID-free system, without introducing any additional loss design. MOTOR reduces the overall space requirements of these models, facilitating information interaction among related items, while also significantly enhancing the model's recommendation capability. Extensive experiments on nine mainstream models demonstrate the significant performance improvement achieved by MOTOR, highlighting its effectiveness in enhancing multimodal recommendation systems.
Large language models (LLMs) have revolutionized natural language processing, albeit at the cost of immense memory and computation requirements. Post-training quantization (PTQ) is becoming the de facto method to reduce the memory footprint and improve the inference throughput of LLMs. In this work, we aim to push the upper limit of LLM PTQ by optimizing the weight rounding parameters with the block reconstruction technique, a predominant method in previous vision models. We propose TesseraQ, a new state-of-the-art PTQ technique, to quantize the weights of LLMs to ultra-low bits. To effectively optimize the rounding in LLMs and stabilize the reconstruction process, we introduce progressive adaptive rounding. This approach iteratively transits the soft rounding variables to hard variables during the reconstruction process. Additionally, we optimize the dequantization scale parameters to fully leverage the block reconstruction technique. We demonstrate that TesseraQ can be seamlessly integrated with existing scaling or clipping-based PTQ algorithms such as AWQ and OmniQuant, significantly enhancing their performance and establishing a new state-of-the-art. For instance, when compared to AWQ, TesseraQ improves the wikitext2 perplexity from 14.65 to 6.82 and average downstream accuracy from 50.52 to 59.27 with 2-bit weight-only quantization of LLaMA-2-7B. Across a range of quantization schemes, including W2A16, W3A16, W3A3, and W4A4, TesseraQ consistently exhibits superior performance.
The recent popular radiance field models, exemplified by Neural Radiance Fields (NeRF), Instant-NGP and 3D Gaussian Splat?ting, are designed to represent 3D content by that training models for each individual scene. This unique characteristic of scene representation and per-scene training distinguishes radiance field models from other neural models, because complex scenes necessitate models with higher representational capacity and vice versa. In this paper, we propose content?aware radiance fields, aligning the model complexity with the scene intricacies through Adversarial Content-Aware Quantization (A-CAQ). Specifically, we make the bitwidth of parameters differentiable and train?able, tailored to the unique characteristics of specific scenes and requirements. The proposed framework has been assessed on Instant-NGP, a well-known NeRF variant and evaluated using various datasets. Experimental results demonstrate a notable reduction in computational complexity, while preserving the requisite reconstruction and rendering quality, making it beneficial for practical deployment of radiance fields models. Codes are available at https://github.com/WeihangLiu2024/Content_Aware_NeRF.
FP8 training has emerged as a promising method for improving training efficiency. Existing frameworks accelerate training by applying FP8 computation to linear layers while leaving optimizer states and activations in higher precision, which fails to fully optimize memory usage. This paper introduces COAT (Compressing Optimizer States and Activations for FP8 Training), a novel FP8 training framework designed to significantly reduce memory footprint when training large models. COAT addresses current limitations through two key innovations: (1) Dynamic Range Expansion, which aligns optimizer state distributions more closely with the FP8 representation range, thereby reducing quantization error, and (2) Mixed-Granularity Activation Quantization, which optimizes activation memory using a combination of per-tensor and per-group quantization strategies. Experiments demonstrate that COAT effectively reduces end-to-end training memory footprint by 1.54x compared to BF16 while achieving nearly lossless performance across various tasks, such as Large Language Model pretraining and fine-tuning and Vision Language Model training. COAT also achieves a 1.43x end-to-end training speedup compared to BF16, performing on par with or surpassing TransformerEngine's speedup. COAT enables efficient full-parameter training of large models on fewer GPUs, and facilitates doubling the batch size in distributed training settings, providing a practical solution for scaling large-scale model training. The code is available at https://github.com/NVlabs/COAT.
Sampling and quantization are standard practices in signal and image processing, but a theoretical understanding of their impact is incomplete. We consider discrete image registration when the underlying function is a one-dimensional spatially-limited piecewise constant function. For ideal noiseless sampling the number of samples from each region of the support of the function generally depends on the placement of the sampling grid. Therefore, if the samples of the function are noisy, then image registration requires alignment and segmentation of the data sequences. One popular strategy for aligning images is selecting the maximum from cross-correlation template matching. To motivate more robust and accurate approaches which also address segmentation, we provide an example of a one-dimensional spatially-limited piecewise constant function for which the cross-correlation technique can perform poorly on noisy samples. While earlier approaches to improve the method involve normalization, our example suggests a novel strategy in our setting. Difference sequences, thresholding, and dynamic programming are well-known techniques in image processing. We prove that they are tools to correctly align and segment noisy data sequences under some conditions on the noise. We also address some of the potential difficulties that could arise in a more general case.
The modulo analog-to-digital converter (ADC) is a promising solution to resolve the limited dynamic range (DR) issue of conventional ADCs and achieve an enhanced digital resolution given a fixed quantization bit budget. However, a modulo ADC requires an unfolding scheme to correct the nonlinear distortion introduced by the modulo operation. This paper presents a sliding discrete Fourier Transform (DFT)-based method for fast signal reconstruction given the modulo ADC output sequence and a 1-bit folding information sequence. In contrast to existing DFT-based signal recovery techniques for modulo ADCs, our proposed sliding DFT method reduces the required observation time and minimizes the spectral leakage effects via proper choice of window function parameters. A mean squared error (MSE) performance guarantee is established for the proposed signal recovery algorithm. More precisely, we derive sufficient conditions for the oversampling factor ($\mathrm{OF}$) and the number of quantization bits ($b$) to obtain a specific MSE performance. Our numerical results demonstrate that modulo ADCs equipped with our proposed recovery method can outperform conventional ADCs without modulo for $\mathrm{OF} \geq 4$ and $b \geq 4$. The impact of spectral leakage on the MSE performance of the proposed sliding DFT recovery method is also quantified.
Text entry with word-gesture keyboards (WGK) is emerging as a popular method and becoming a key interaction for Extended Reality (XR). However, the diversity of interaction modes, keyboard sizes, and visual feedback in these environments introduces divergent word-gesture trajectory data patterns, thus leading to complexity in decoding trajectories into text. Template-matching decoding methods, such as SHARK^2, are commonly used for these WGK systems because they are easy to implement and configure. However, these methods are susceptible to decoding inaccuracies for noisy trajectories. While conventional neural-network-based decoders (neural decoders) trained on word-gesture trajectory data have been proposed to improve accuracy, they have their own limitations: they require extensive data for training and deep-learning expertise for implementation. To address these challenges, we propose a novel solution that combines ease of implementation with high decoding accuracy: a generalizable neural decoder enabled by pre-training on large-scale coarsely discretized word-gesture trajectories. This approach produces a ready-to-use WGK decoder that is generalizable across mid-air and on-surface WGK systems in augmented reality (AR) and virtual reality (VR), which is evident by a robust average Top-4 accuracy of 90.4% on four diverse datasets. It significantly outperforms SHARK^2 with a 37.2% enhancement and surpasses the conventional neural decoder by 7.4%. Moreover, the Pre-trained Neural Decoder's size is only 4 MB after quantization, without sacrificing accuracy, and it can operate in real-time, executing in just 97 milliseconds on Quest 3.