Mixed-precision computing, a widely applied technique in AI, offers a larger trade-off space between accuracy and efficiency. The recent purposed Mixed-Precision Over-the-Air Federated Learning (MP-OTA-FL) enables clients to operate at appropriate precision levels based on their heterogeneous hardware, taking advantages of the larger trade-off space while covering the quantization overheads in the mixed-precision modulation scheme for the OTA aggregation process. A key to further exploring the potential of the MP-OTA-FL framework is the optimization of client precision levels. The choice of precision level hinges on multifaceted factors including hardware capability, potential client contribution, and user satisfaction, among which factors can be difficult to define or quantify. In this paper, we propose a RAG-based User Profiling for precision planning framework that integrates retrieval-augmented LLMs and dynamic client profiling to optimize satisfaction and contributions. This includes a hybrid interface for gathering device/user insights and an RAG database storing historical quantization decisions with feedback. Experiments show that our method boosts satisfaction, energy savings, and global model accuracy in MP-OTA-FL systems.
Model quantization enables efficient deployment of deep neural networks on edge devices through low-bit parameter representation, yet raises critical challenges for implementing machine unlearning (MU) under data privacy regulations. Existing MU methods designed for full-precision models fail to address two fundamental limitations in quantized networks: 1) Noise amplification from label mismatch during data processing, and 2) Gradient imbalance between forgotten and retained data during training. These issues are exacerbated by quantized models' constrained parameter space and discrete optimization. We propose Q-MUL, the first dedicated unlearning framework for quantized models. Our method introduces two key innovations: 1) Similar Labels assignment replaces random labels with semantically consistent alternatives to minimize noise injection, and 2) Adaptive Gradient Reweighting dynamically aligns parameter update contributions from forgotten and retained data. Through systematic analysis of quantized model vulnerabilities, we establish theoretical foundations for these mechanisms. Extensive evaluations on benchmark datasets demonstrate Q-MUL's superiority over existing approaches.
Speculative decoding (SD) has emerged as a method to accelerate LLM inference without sacrificing any accuracy over the 16-bit model inference. In a typical SD setup, the idea is to use a full-precision, small, fast model as "draft" to generate the next few tokens and use the "target" large model to verify the draft-generated tokens. The efficacy of this method heavily relies on the acceptance ratio of the draft-generated tokens and the relative token throughput of the draft versus the target model. Nevertheless, an efficient SD pipeline requires pre-training and aligning the draft model to the target model, making it impractical for LLM inference in a plug-and-play fashion. In this work, we propose using MXFP4 models as drafts in a plug-and-play fashion since the MXFP4 Weight-Only-Quantization (WOQ) merely direct-casts the BF16 target model weights to MXFP4. In practice, our plug-and-play solution gives speedups up to 2x over the BF16 baseline. Then we pursue an opportunity for further acceleration: the MXFP4 draft token generation itself can be accelerated via speculative decoding by using yet another smaller draft. We call our method ML-SpecQD: Multi-Level Speculative Decoding with Quantized Drafts since it recursively applies speculation for accelerating the draft-token generation. Combining Multi-Level Speculative Decoding with MXFP4 Quantized Drafts we outperform state-of-the-art speculative decoding, yielding speedups up to 2.72x over the BF16 baseline.
Current transformer-based imitation learning approaches introduce discrete action representations and train an autoregressive transformer decoder on the resulting latent code. However, the initial quantization breaks the continuous structure of the action space thereby limiting the capabilities of the generative model. We propose a quantization-free method instead that leverages Generative Infinite-Vocabulary Transformers (GIVT) as a direct, continuous policy parametrization for autoregressive transformers. This simplifies the imitation learning pipeline while achieving state-of-the-art performance on a variety of popular simulated robotics tasks. We enhance our policy roll-outs by carefully studying sampling algorithms, further improving the results.
This work focuses on full-body co-speech gesture generation. Existing methods typically employ an autoregressive model accompanied by vector-quantized tokens for gesture generation, which results in information loss and compromises the realism of the generated gestures. To address this, inspired by the natural continuity of real-world human motion, we propose MAG, a novel multi-modal aligned framework for high-quality and diverse co-speech gesture synthesis without relying on discrete tokenization. Specifically, (1) we introduce a motion-text-audio-aligned variational autoencoder (MTA-VAE), which leverages pre-trained WavCaps' text and audio embeddings to enhance both semantic and rhythmic alignment with motion, ultimately producing more realistic gestures. (2) Building on this, we propose a multimodal masked autoregressive model (MMAG) that enables autoregressive modeling in continuous motion embeddings through diffusion without vector quantization. To further ensure multi-modal consistency, MMAG incorporates a hybrid granularity audio-text fusion block, which serves as conditioning for diffusion process. Extensive experiments on two benchmark datasets demonstrate that MAG achieves stateof-the-art performance both quantitatively and qualitatively, producing highly realistic and diverse co-speech gestures.The code will be released to facilitate future research.
Effective training and debriefing are critical in high-stakes, mission-critical environments such as disaster response, military simulations, and industrial safety, where precision and minimizing errors are paramount. The traditional post-training analysis relies on manually reviewing 2D videos, a time-consuming process that lacks comprehensive situational awareness. To address these limitations, we introduce ACT360, a system that leverages 360-degree videos and machine learning for automated action detection and structured debriefing. ACT360 integrates 360YOWO, an enhanced You Only Watch Once (YOWO) model with spatial attention and equirectangular-aware convolution (EAC) to mitigate panoramic video distortions. To enable deployment in resource-constrained environments, we apply quantization and model pruning, reducing the model size by 74% while maintaining robust accuracy (mAP drop of only 1.5%, from 0.865 to 0.850) and improving inference speed. We validate our approach on a publicly available dataset of 55 labeled 360-degree videos covering seven key operational actions, recorded across various real-world training sessions and environmental conditions. Additionally, ACT360 integrates 360AIE (Action Insight Explorer), a web-based interface for automatic action detection, retrieval, and textual summarization using large language models (LLMs), significantly enhancing post-incident analysis efficiency. ACT360 serves as a generalized framework for mission-critical debriefing, incorporating EAC, spatial attention, summarization, and model optimization. These innovations apply to any training environment requiring lightweight action detection and structured post-exercise analysis.
We present a versatile latent representation that enables physically simulated character to efficiently utilize motion priors. To build a powerful motion embedding that is shared across multiple tasks, the physics controller should employ rich latent space that is easily explored and capable of generating high-quality motion. We propose integrating continuous and discrete latent representations to build a versatile motion prior that can be adapted to a wide range of challenging control tasks. Specifically, we build a discrete latent model to capture distinctive posterior distribution without collapse, and simultaneously augment the sampled vector with the continuous residuals to generate high-quality, smooth motion without jittering. We further incorporate Residual Vector Quantization, which not only maximizes the capacity of the discrete motion prior, but also efficiently abstracts the action space during the task learning phase. We demonstrate that our agent can produce diverse yet smooth motions simply by traversing the learned motion prior through unconditional motion generation. Furthermore, our model robustly satisfies sparse goal conditions with highly expressive natural motions, including head-mounted device tracking and motion in-betweening at irregular intervals, which could not be achieved with existing latent representations.
Quantization-Aware Training (QAT) is one of the prevailing neural network compression solutions. However, its stability has been challenged for yielding deteriorating performances as the quantization error is inevitable. We find that the sharp landscape of loss, which leads to a dramatic performance drop, is an essential factor that causes instability. Theoretically, we have discovered that the perturbations in the feature would bring a flat local minima. However, simply adding perturbations into either weight or feature empirically deteriorates the performance of the Full Precision (FP) model. In this paper, we propose Feature-Perturbed Quantization (FPQ) to stochastically perturb the feature and employ the feature distillation method to the quantized model. Our method generalizes well to different network architectures and various QAT methods. Furthermore, we mathematically show that FPQ implicitly regularizes the Hessian norm, which calibrates the smoothness of a loss landscape. Extensive experiments demonstrate that our approach significantly outperforms the current State-Of-The-Art (SOTA) QAT methods and even the FP counterparts.
Flat minima, known to enhance generalization and robustness in supervised learning, remain largely unexplored in generative models. In this work, we systematically investigate the role of loss surface flatness in generative models, both theoretically and empirically, with a particular focus on diffusion models. We establish a theoretical claim that flatter minima improve robustness against perturbations in target prior distributions, leading to benefits such as reduced exposure bias -- where errors in noise estimation accumulate over iterations -- and significantly improved resilience to model quantization, preserving generative performance even under strong quantization constraints. We further observe that Sharpness-Aware Minimization (SAM), which explicitly controls the degree of flatness, effectively enhances flatness in diffusion models, whereas other well-known methods such as Stochastic Weight Averaging (SWA) and Exponential Moving Average (EMA), which promote flatness indirectly via ensembling, are less effective. Through extensive experiments on CIFAR-10, LSUN Tower, and FFHQ, we demonstrate that flat minima in diffusion models indeed improves not only generative performance but also robustness.
We present OuroMamba, the first data-free post-training quantization (DFQ) method for vision Mamba-based models (VMMs). We identify two key challenges in enabling DFQ for VMMs, (1) VMM's recurrent state transitions restricts capturing of long-range interactions and leads to semantically weak synthetic data, (2) VMM activations exhibit dynamic outlier variations across time-steps, rendering existing static PTQ techniques ineffective. To address these challenges, OuroMamba presents a two-stage framework: (1) OuroMamba-Gen to generate semantically rich and meaningful synthetic data. It applies contrastive learning on patch level VMM features generated through neighborhood interactions in the latent state space, (2) OuroMamba-Quant to employ mixed-precision quantization with lightweight dynamic outlier detection during inference. In specific, we present a thresholding based outlier channel selection strategy for activations that gets updated every time-step. Extensive experiments across vision and generative tasks show that our data-free OuroMamba surpasses existing data-driven PTQ techniques, achieving state-of-the-art performance across diverse quantization settings. Additionally, we implement efficient GPU kernels to achieve practical latency speedup of up to 2.36x. Code will be released soon.
Tomato maturity plays a pivotal role in optimizing harvest timing and ensuring product quality, but current methods struggle to achieve high accuracy along computational efficiency simultaneously. Existing deep learning approaches, while accurate, are often too computationally demanding for practical use in resource-constrained agricultural settings. In contrast, simpler techniques fail to capture the nuanced features needed for precise classification. This study aims to develop a computationally efficient tomato classification model using the ResNet-18 architecture optimized through transfer learning, pruning, and quantization techniques. Our objective is to address the dual challenge of maintaining high accuracy while enabling real-time performance on low-power edge devices. Then, these models were deployed on an edge device to investigate their performance for tomato maturity classification. The quantized model achieved an accuracy of 97.81%, with an average classification time of 0.000975 seconds per image. The pruned and auto-tuned model also demonstrated significant improvements in deployment metrics, further highlighting the benefits of optimization techniques. These results underscore the potential for a balanced solution that meets the accuracy and efficiency demands of modern agricultural production, paving the way for practical, real-world deployment in resource-limited environments.
Vector Quantization (VQ) techniques face significant challenges in codebook utilization, limiting reconstruction fidelity in image modeling. We introduce a Dual Codebook mechanism that effectively addresses this limitation by partitioning the representation into complementary global and local components. The global codebook employs a lightweight transformer for concurrent updates of all code vectors, while the local codebook maintains precise feature representation through deterministic selection. This complementary approach is trained from scratch without requiring pre-trained knowledge. Experimental evaluation across multiple standard benchmark datasets demonstrates state-of-the-art reconstruction quality while using a compact codebook of size 512 - half the size of previous methods that require pre-training. Our approach achieves significant FID improvements across diverse image domains, particularly excelling in scene and face reconstruction tasks. These results establish Dual Codebook VQ as an efficient paradigm for high-fidelity image reconstruction with significantly reduced computational requirements.
The growing volume of high-resolution Whole Slide Images in digital histopathology poses significant storage, transmission, and computational efficiency challenges. Standard compression methods, such as JPEG, reduce file sizes but often fail to preserve fine-grained phenotypic details critical for downstream tasks. In this work, we repurpose autoencoders (AEs) designed for Latent Diffusion Models as an efficient learned compression framework for pathology images. We systematically benchmark three AE models with varying compression levels and evaluate their reconstruction ability using pathology foundation models. We introduce a fine-tuning strategy to further enhance reconstruction fidelity that optimizes a pathology-specific learned perceptual metric. We validate our approach on downstream tasks, including segmentation, patch classification, and multiple instance learning, showing that replacing images with AE-compressed reconstructions leads to minimal performance degradation. Additionally, we propose a K-means clustering-based quantization method for AE latents, improving storage efficiency while maintaining reconstruction quality. We provide the weights of the fine-tuned autoencoders at https://huggingface.co/collections/StonyBrook-CVLab/pathology-fine-tuned-aes-67d45f223a659ff2e3402dd0.
Post-training pruning, as one of the key techniques for compressing large language models, plays a vital role in lightweight model deployment and model sparsity. However, current mainstream pruning methods dependent on the Hessian matrix face significant limitations in both pruning speed and practical effectiveness due to the computationally intensive nature of second-order derivative calculations. This paper presents SwiftPrune, a novel Hessian-free weight pruning method that achieves hardware-efficient model compression through two key innovations: 1) SwiftPrune eliminates the need for computationally intensive Hessian matrix calculations by introducing a contribution-based weight metric, which evaluates the importance of weights without relying on second-order derivatives. 2) we employ the Exponentially Weighted Moving Average (EWMA) technique to bypass weight sorting, enabling the selection of weights that contribute most to LLM accuracy and further reducing time complexity. Our approach is extended to support structured sparsity pruning, facilitating efficient execution on modern hardware accelerators. We validate the SwiftPrune on three LLMs (namely LLaMA2, LLaMA3, and Pythia), demonstrating that it significantly enhances compression performance. The experimental findings reveal that SwiftPrune completes the pruning process within seconds, achieving an average speedup of 12.29x (up to 56.02x) over existing SOTA approaches.
The paper addresses the synchronization of multi-agent systems with continuous-time dynamics interacting through a very general class of monotonic continuous signal functions that covers estimation biases, approximation of discrete quantization, or state-dependent estimation. Our analysis reveals that, in the setup under consideration, synchronization equilibria are exactly the fixed points of the signal function. We also derive intuitive stability conditions based on whether the signal underestimates or overestimates the state of the agents around these fixed points. Moreover, we show that network topology plays a crucial role in asymptotic synchronization. These results provide interesting insights into the interplay between communication nonlinearity and network connectivity, paving the way for advanced coordination strategies in complex systems.
Automated speech recognition (ASR) models have gained prominence for applications such as captioning, speech translation, and live transcription. This paper studies Whisper and two model variants: one optimized for live speech streaming and another for offline transcription. Notably, these models have been found to generate hallucinated content, reducing transcription reliability. Furthermore, larger model variants exhibit increased latency and pose challenges for deployment on resource-constrained devices. This study analyzes the similarities and differences between three Whisper models, qualitatively examining their distinct capabilities. Next, this study quantifies the impact of model quantization on latency and evaluates its viability for edge deployment. Using the open source LibriSpeech dataset, this paper evaluates the word error rate (WER) along with latency analysis of whispercpp using 3 quantization methods (INT4, INT5, INT8). Results show that quantization reduces latency by 19\% and model size by 45\%, while preserving transcription accuracy. These findings provide insights into the optimal use cases of different Whisper models and edge device deployment possibilities. All code, datasets, and implementation details are available in a public GitHub repository: https://github.com/allisonandreyev/WhisperQuantization.git
3D Gaussian Splatting (3DGS) achieves impressive rendering fidelity and speed for novel view synthesis. However, its substantial data size poses a significant challenge for practical applications. While many compression techniques have been proposed, they fail to efficiently utilize existing bitstreams in on-demand applications due to their lack of progressivity, leading to a waste of resource. To address this issue, we propose PCGS (Progressive Compression of 3D Gaussian Splatting), which adaptively controls both the quantity and quality of Gaussians (or anchors) to enable effective progressivity for on-demand applications. Specifically, for quantity, we introduce a progressive masking strategy that incrementally incorporates new anchors while refining existing ones to enhance fidelity. For quality, we propose a progressive quantization approach that gradually reduces quantization step sizes to achieve finer modeling of Gaussian attributes. Furthermore, to compact the incremental bitstreams, we leverage existing quantization results to refine probability prediction, improving entropy coding efficiency across progressive levels. Overall, PCGS achieves progressivity while maintaining compression performance comparable to SoTA non-progressive methods. Code available at: github.com/YihangChen-ee/PCGS.
Transformer models have achieved remarkable success across various AI applications but face significant training costs. Low-bit training, such as INT8 training, can leverage computational units with higher throughput, and has already demonstrated its effectiveness on GPT2 models with block-level quantization. However, it struggles with modern Transformer variants incorporating GLU units. This is because those variants demonstrate complex distributions of activation outliers. To address the challenge, we propose Fallback Quantization, implementing mixed-precision GEMM that dynamically falls back 8-bit to 16-bit for activation blocks containing outliers. Experiments show that our approach is robustly competent in both fine-tuning and pretraining settings. Moreover, our method achieves a 1.57x end-to-end training speedup on RTX4090 GPUs.
Training large-scale neural networks in vision, and multimodal domains demands substantial memory resources, primarily due to the storage of optimizer states. While LoRA, a popular parameter-efficient method, reduces memory usage, it often suffers from suboptimal performance due to the constraints of low-rank updates. Low-rank gradient projection methods (e.g., GaLore, Flora) reduce optimizer memory by projecting gradients and moment estimates into low-rank spaces via singular value decomposition or random projection. However, they fail to account for inter-projection correlation, causing performance degradation, and their projection strategies often incur high computational costs. In this paper, we present COAP (Correlation-Aware Gradient Projection), a memory-efficient method that minimizes computational overhead while maintaining training performance. Evaluated across various vision, language, and multimodal tasks, COAP outperforms existing methods in both training speed and model performance. For LLaMA-1B, it reduces optimizer memory by 61% with only 2% additional time cost, achieving the same PPL as AdamW. With 8-bit quantization, COAP cuts optimizer memory by 81% and achieves 4x speedup over GaLore for LLaVA-v1.5-7B fine-tuning, while delivering higher accuracy.
Visual Mamba networks (ViMs) extend the selective space state model (Mamba) to various vision tasks and demonstrate significant potential. Vector quantization (VQ), on the other hand, decomposes network weights into codebooks and assignments, significantly reducing memory usage and computational latency to enable ViMs deployment on edge devices. Although existing VQ methods have achieved extremely low-bit quantization (e.g., 3-bit, 2-bit, and 1-bit) in convolutional neural networks and Transformer-based networks, directly applying these methods to ViMs results in unsatisfactory accuracy. We identify several key challenges: 1) The weights of Mamba-based blocks in ViMs contain numerous outliers, significantly amplifying quantization errors. 2) When applied to ViMs, the latest VQ methods suffer from excessive memory consumption, lengthy calibration procedures, and suboptimal performance in the search for optimal codewords. In this paper, we propose ViM-VQ, an efficient post-training vector quantization method tailored for ViMs. ViM-VQ consists of two innovative components: 1) a fast convex combination optimization algorithm that efficiently updates both the convex combinations and the convex hulls to search for optimal codewords, and 2) an incremental vector quantization strategy that incrementally confirms optimal codewords to mitigate truncation errors. Experimental results demonstrate that ViM-VQ achieves state-of-the-art performance in low-bit quantization across various visual tasks.