Contemporary machine learning models, such as language models, are powerful, but come with immense resource requirements both at training and inference time. It has been shown that decoder-only language models can be trained to a competitive state with ternary weights (1.58 bits per weight), facilitating efficient inference. Here, we start our exploration with non-transformer model architectures, investigating 1.58-bit training for multi-layer perceptrons and graph neural networks. Then, we explore 1.58-bit training in other transformer-based language models, namely encoder-only and encoder-decoder models. Our results show that in all of these settings, 1.58-bit training is on par with or sometimes even better than the standard 32/16-bit models.
This study focuses on Intelligent Fault Diagnosis (IFD) in rotating machinery utilizing a single microphone and a data-driven methodology, effectively diagnosing 42 classes of fault types and severities. The research leverages sound data from the imbalanced MaFaulDa dataset, aiming to strike a balance between high performance and low resource consumption. The testing phase encompassed a variety of configurations, including sampling, quantization, signal normalization, silence removal, Wiener filtering, data scaling, windowing, augmentation, and classifier tuning using XGBoost. Through the analysis of time, frequency, mel-frequency, and statistical features, we achieved an impressive accuracy of 99.54% and an F-Beta score of 99.52% with just 6 boosting trees at an 8 kHz, 8-bit configuration. Moreover, when utilizing only MFCCs along with their first- and second-order deltas, we recorded an accuracy of 97.83% and an F-Beta score of 97.67%. Lastly, by implementing a greedy wrapper approach, we obtained a remarkable accuracy of 96.82% and an F-Beta score of 98.86% using 50 selected features, nearly all of which were first- and second-order deltas of the MFCCs.
This paper presents a comprehensive analysis of quantization techniques for optimizing Large Language Models (LLMs), specifically focusing on Post-Training Quantization (PTQ) and Quantization-Aware Training (QAT). Through empirical evaluation across models ranging from 10M to 1B parameters, we demonstrate that quantization can achieve up to 68% reduction in model size while maintaining performance within 6% of full-precision baselines when utilizing our proposed scaling factor {\gamma}. Our experiments show that INT8 quantization delivers a 40% reduction in computational cost and power consumption, while INT4 quantization further improves these metrics by 60%. We introduce a novel theoretical framework for mixed-precision quantization, deriving optimal bit allocation strategies based on layer sensitivity and weight variance. Hardware efficiency evaluations on edge devices reveal that our quantization approach enables up to 2.4x throughput improvement for INT8 and 3x for INT4, with 60% power reduction compared to full-precision models.
Combining gradient compression methods (e.g., CountSketch, quantization) and adaptive optimizers (e.g., Adam, AMSGrad) is a desirable goal in federated learning (FL), with potential benefits on both fewer communication rounds and less per-round communication. In spite of the preliminary empirical success of sketched adaptive methods, existing convergence analyses show the communication cost to have a linear dependence on the ambient dimension, i.e., number of parameters, which is prohibitively high for modern deep learning models. In this work, we introduce specific sketched adaptive federated learning (SAFL) algorithms and, as our main contribution, provide theoretical convergence analyses in different FL settings with guarantees on communication cost depending only logarithmically (instead of linearly) on the ambient dimension. Unlike existing analyses, we show that the entry-wise sketching noise existent in the preconditioners and the first moments of SAFL can be implicitly addressed by leveraging the recently-popularized anisotropic curvatures in deep learning losses, e.g., fast decaying loss Hessian eigen-values. In the i.i.d. client setting of FL, we show that SAFL achieves asymptotic $O(1/\sqrt{T})$ convergence, and converges faster in the initial epochs. In the non-i.i.d. client setting, where non-adaptive methods lack convergence guarantees, we show that SACFL (SAFL with clipping) algorithms can provably converge in spite of the additional heavy-tailed noise. Our theoretical claims are supported by empirical studies on vision and language tasks, and in both fine-tuning and training-from-scratch regimes. Surprisingly, as a by-product of our analysis, the proposed SAFL methods are competitive with the state-of-the-art communication-efficient federated learning algorithms based on error feedback.
Text emotion detection constitutes a crucial foundation for advancing artificial intelligence from basic comprehension to the exploration of emotional reasoning. Most existing emotion detection datasets rely on manual annotations, which are associated with high costs, substantial subjectivity, and severe label imbalances. This is particularly evident in the inadequate annotation of micro-emotions and the absence of emotional intensity representation, which fail to capture the rich emotions embedded in sentences and adversely affect the quality of downstream task completion. By proposing an all-labels and training-set label regression method, we map label values to energy intensity levels, thereby fully leveraging the learning capabilities of machine models and the interdependencies among labels to uncover multiple emotions within samples. This led to the establishment of the Emotion Quantization Network (EQN) framework for micro-emotion detection and annotation. Using five commonly employed sentiment datasets, we conducted comparative experiments with various models, validating the broad applicability of our framework within NLP machine learning models. Based on the EQN framework, emotion detection and annotation are conducted on the GoEmotions dataset. A comprehensive comparison with the results from Google literature demonstrates that the EQN framework possesses a high capability for automatic detection and annotation of micro-emotions. The EQN framework is the first to achieve automatic micro-emotion annotation with energy-level scores, providing strong support for further emotion detection analysis and the quantitative research of emotion computing.
Deep learning methods have established a significant place in image classification. While prior research has focused on enhancing final outcomes, the opaque nature of the decision-making process in these models remains a concern for experts. Additionally, the deployment of these methods can be problematic in resource-limited environments. This paper tackles the inherent black-box nature of these models by providing real-time explanations during the training phase, compelling the model to concentrate on the most distinctive and crucial aspects of the input. Furthermore, we employ established quantization techniques to address resource constraints. To assess the effectiveness of our approach, we explore how quantization influences the interpretability and accuracy of Convolutional Neural Networks through a comparative analysis of saliency maps from standard and quantized models. Quantization is implemented during the training phase using the Parameterized Clipping Activation method, with a focus on the MNIST and FashionMNIST benchmark datasets. We evaluated three bit-width configurations (2-bit, 4-bit, and mixed 4/2-bit) to explore the trade-off between efficiency and interpretability, with each configuration designed to highlight varying impacts on saliency map clarity and model accuracy. The results indicate that while quantization is crucial for implementing models on resource-limited devices, it necessitates a trade-off between accuracy and interpretability. Lower bit-widths result in more pronounced reductions in both metrics, highlighting the necessity of meticulous quantization parameter selection in applications where model transparency is paramount. The study underscores the importance of achieving a balance between efficiency and interpretability in the deployment of neural networks.
Vision Language Models (VLMs) are central to Visual Question Answering (VQA) systems and are typically deployed in the cloud due to their high computational demands. However, this cloud-only approach underutilizes edge computational resources and requires significant bandwidth for transmitting raw images. In this paper, we introduce an edge-cloud collaborative VQA system, called LLaVA-AlignedVQ, which features a novel Aligned Vector Quantization algorithm (AlignedVQ) that efficiently compress intermediate features without compromising accuracy to support partitioned execution. Our experiments demonstrate that LLaVA-AlignedVQ achieves approximately 1365x compression rate of intermediate features, reducing data transmission overhead by 96.8% compared to transmitting JPEG90-compressed images to the cloud. LLaVA-AlignedVQ achieves an inference speedup of 2-15x while maintaining high accuracy, remaining within -2.23% to +1.6% of the original model's accuracy performance across eight VQA datasets, compared to the cloud-only solution.
We present an innovative approach for solving mathematical problems in Bengali, developed for the DL Sprint 3.0 BUET CSE Fest 2024 Competition. Our method uses advanced deep learning models, notably the Qwen 2.5 series, with improvements made through prompt engineering, model quantization, and Tool Integrated Reasoning (TIR) to handle complex calculations. Initially, we explored various model architectures, including fine-tuned Mistral and quantized Qwen models, refining them with translation techniques, Retrieval-Augmented Generation (RAG), and custom dataset curation. Manual hyperparameter tuning optimized parameters like temperature and top-p to enhance model adaptability and accuracy. Removal of RAG and parameter adjustments further improved robustness. Our approach highlights the potential of advanced NLP techniques in solving Bengali mathematical problems.
Federated fine-tuning of pre-trained Large Language Models (LLMs) enables task-specific adaptation across diverse datasets while preserving privacy. However, challenges such as high computational and memory demands, heterogeneous client resources, bandwidth constraints, and ineffective global aggregation hinder its efficiency. To address these issues, we propose HAFLQ (Heterogeneous Adaptive Federated Low-Rank Adaptation Fine-tuned LLM with Quantization), a novel framework for efficient and scalable federated fine-tuning of LLMs in heterogeneous environments. To reduce memory and computation demands, we propose a salience-driven adaptive LLM quantization framework that evaluates the importance of transformer blocks using a salience metric and applies adaptive block-wise quantization accordingly. To handle heterogeneous computational capabilities, we propose an importance-based parameter truncation and freezing scheme. To address communication bottlenecks, we propose an importance-aware bandwidth-adaptive quantization method, which dynamically adjusts parameter precision based on importance and bandwidth constraints. To improve global model aggregation, we propose an adaptive rank-1 matrix-level aggregation strategy, which prevents information dilution and accelerates convergence by aggregating only updated rank-1 matrices from clients. Experimental results on the text classification task demonstrate that HAFLQ reduces memory usage by 31%, lowers communication cost by 49%, improves accuracy by 50%, and achieves faster convergence compared to the baseline method.
Recent works have shown a surprising result: a small fraction of Large Language Model (LLM) parameter outliers are disproportionately important to the quality of the model. LLMs contain billions of parameters, so these small fractions, such as 0.01%, translate to hundreds of thousands of parameters. In this work, we present an even more surprising finding: Pruning as few as a single parameter can destroy an LLM's ability to generate text -- increasing perplexity by 3 orders of magnitude and reducing zero-shot accuracy to guessing. We propose a data-free method for identifying such parameters, termed super weights, using a single forward pass through the model. We additionally find that these super weights induce correspondingly rare and large activation outliers, termed super activations. When preserved with high precision, super activations can improve simple round-to-nearest quantization to become competitive with state-of-the-art methods. For weight quantization, we similarly find that by preserving the super weight and clipping other weight outliers, round-to-nearest quantization can scale to much larger block sizes than previously considered. To facilitate further research into super weights, we provide an index of super weight coordinates for common, openly available LLMs.
The neural radiance fields (NeRF) have advanced the development of 3D volumetric video technology, but the large data volumes they involve pose significant challenges for storage and transmission. To address these problems, the existing solutions typically compress these NeRF representations after the training stage, leading to a separation between representation training and compression. In this paper, we try to directly learn a compact NeRF representation for volumetric video in the training stage based on the proposed rate-aware compression framework. Specifically, for volumetric video, we use a simple yet effective modeling strategy to reduce temporal redundancy for the NeRF representation. Then, during the training phase, an implicit entropy model is utilized to estimate the bitrate of the NeRF representation. This entropy model is then encoded into the bitstream to assist in the decoding of the NeRF representation. This approach enables precise bitrate estimation, thereby leading to a compact NeRF representation. Furthermore, we propose an adaptive quantization strategy and learn the optimal quantization step for the NeRF representations. Finally, the NeRF representation can be optimized by using the rate-distortion trade-off. Our proposed compression framework can be used for different representations and experimental results demonstrate that our approach significantly reduces the storage size with marginal distortion and achieves state-of-the-art rate-distortion performance for volumetric video on the HumanRF and ReRF datasets. Compared to the previous state-of-the-art method TeTriRF, we achieved an approximately -80% BD-rate on the HumanRF dataset and -60% BD-rate on the ReRF dataset.
Quantization of foundational models (FMs) is significantly more challenging than traditional DNNs due to the emergence of large magnitude values called outliers. Existing outlier-aware algorithm-architecture co-design techniques either use mixed-precision, retaining outliers at high precision but compromise hardware efficiency, or quantize inliers and outliers at the same precision, improving hardware efficiency at the cost of accuracy. To address this mutual exclusivity, we propose MicroScopiQ, a novel co-design technique that leverages pruning to complement outlier-aware quantization. MicroScopiQ retains outliers at higher precision while pruning a certain fraction of least important weights to distribute the additional outlier bits; ensuring high accuracy, aligned memory and hardware efficiency. We design a high-throughput, low overhead accelerator architecture composed of multi-precision INT processing elements and a network-on-chip called ReCoN that efficiently abstracts the complexity of supporting high-precision outliers. Additionally, unlike prior techniques, MicroScopiQ does not assume any locality of outlier weights, enabling applicability to a broad range of FMs. Extensive experiments across diverse quantization settings demonstrate that MicroScopiQ achieves state-of-the-art quantization accuracy, while delivering up to 3x faster inference and 2x lower energy consumption compared to existing alternatives. Code is available at: https://github.com/georgia-tech-synergy-lab/MicroScopiQ-LLM-Quantization
Federated Learning has emerged as a leading approach for decentralized machine learning, enabling multiple clients to collaboratively train a shared model without exchanging private data. While FL enhances data privacy, it remains vulnerable to inference attacks, such as gradient inversion and membership inference, during both training and inference phases. Homomorphic Encryption provides a promising solution by encrypting model updates to protect against such attacks, but it introduces substantial communication overhead, slowing down training and increasing computational costs. To address these challenges, we propose QuanCrypt-FL, a novel algorithm that combines low-bit quantization and pruning techniques to enhance protection against attacks while significantly reducing computational costs during training. Further, we propose and implement mean-based clipping to mitigate quantization overflow or errors. By integrating these methods, QuanCrypt-FL creates a communication-efficient FL framework that ensures privacy protection with minimal impact on model accuracy, thereby improving both computational efficiency and attack resilience. We validate our approach on MNIST, CIFAR-10, and CIFAR-100 datasets, demonstrating superior performance compared to state-of-the-art methods. QuanCrypt-FL consistently outperforms existing method and matches Vanilla-FL in terms of accuracy across varying client. Further, QuanCrypt-FL achieves up to 9x faster encryption, 16x faster decryption, and 1.5x faster inference compared to BatchCrypt, with training time reduced by up to 3x.
Efficient wideband spectrum sensing (WSS) is essential for managing spectrum scarcity in wireless communications. However, existing compressed sensing (CS)-based WSS methods require high sampling rates and power consumption, particularly with high-precision analog-to-digital converters (ADCs). Although 1-bit CS with low-precision ADCs can mitigate these demands, most approaches still depend on multi-user cooperation and prior sparsity information, which are often unavailable in WSS scenarios. This paper introduces a non-cooperative WSS method using multicoset sampling with 1-bit ADCs to achieve sub-Nyquist sampling without requiring sparsity knowledge. We analyze the impact of 1-bit quantization on multiband signals, then apply eigenvalue decomposition to isolate the signal subspace from noise, enabling spectrum support estimation without signal reconstruction. This approach provides a power-efficient solution for WSS that eliminates the need for cooperation and prior information.
Recent research on the 1-bit Large Language Models (LLMs), such as BitNet b1.58, presents a promising direction for reducing the inference cost of LLMs while maintaining their performance. In this work, we introduce BitNet a4.8, enabling 4-bit activations for 1-bit LLMs. BitNet a4.8 employs a hybrid quantization and sparsification strategy to mitigate the quantization errors introduced by the outlier channels. Specifically, we utilize 4-bit activations for inputs to the attention and feed-forward network layers, while sparsifying intermediate states followed with 8-bit quantization. Extensive experiments demonstrate that BitNet a4.8 achieves performance comparable to BitNet b1.58 with equivalent training costs, while being faster in inference with enabling 4-bit (INT4/FP4) kernels. Additionally, BitNet a4.8 activates only 55% of parameters and supports 3-bit KV cache, further enhancing the efficiency of large-scale LLM deployment and inference.
Low precision training and inference affect both the quality and cost of language models, but current scaling laws do not account for this. In this work, we devise "precision-aware" scaling laws for both training and inference. We propose that training in lower precision reduces the model's "effective parameter count," allowing us to predict the additional loss incurred from training in low precision and post-train quantization. For inference, we find that the degradation introduced by post-training quantization increases as models are trained on more data, eventually making additional pretraining data actively harmful. For training, our scaling laws allow us to predict the loss of a model with different parts in different precisions, and suggest that training larger models in lower precision may be compute optimal. We unify the scaling laws for post and pretraining quantization to arrive at a single functional form that predicts degradation from training and inference in varied precisions. We fit on over 465 pretraining runs and validate our predictions on model sizes up to 1.7B parameters trained on up to 26B tokens.
Diffusion models have been proven highly effective at generating high-quality images. However, as these models grow larger, they require significantly more memory and suffer from higher latency, posing substantial challenges for deployment. In this work, we aim to accelerate diffusion models by quantizing their weights and activations to 4 bits. At such an aggressive level, both weights and activations are highly sensitive, where conventional post-training quantization methods for large language models like smoothing become insufficient. To overcome this limitation, we propose SVDQuant, a new 4-bit quantization paradigm. Different from smoothing which redistributes outliers between weights and activations, our approach absorbs these outliers using a low-rank branch. We first consolidate the outliers by shifting them from activations to weights, then employ a high-precision low-rank branch to take in the weight outliers with Singular Value Decomposition (SVD). This process eases the quantization on both sides. However, na\"{\i}vely running the low-rank branch independently incurs significant overhead due to extra data movement of activations, negating the quantization speedup. To address this, we co-design an inference engine Nunchaku that fuses the kernels of the low-rank branch into those of the low-bit branch to cut off redundant memory access. It can also seamlessly support off-the-shelf low-rank adapters (LoRAs) without the need for re-quantization. Extensive experiments on SDXL, PixArt-$\Sigma$, and FLUX.1 validate the effectiveness of SVDQuant in preserving image quality. We reduce the memory usage for the 12B FLUX.1 models by 3.5$\times$, achieving 3.0$\times$ speedup over the 4-bit weight-only quantized baseline on the 16GB laptop 4090 GPU, paving the way for more interactive applications on PCs. Our quantization library and inference engine are open-sourced.
We consider a distributed detection problem within a wireless sensor network (WSN), where a substantial number of sensors cooperate to detect the existence of sparse stochastic signals. To achieve a trade-off between detection performance and system constraints, multi-bit quantizers are employed at local sensors. Then, two quantization strategies, namely raw quantization (RQ) and likelihood ratio quantization (LQ), are examined. The multi-bit quantized signals undergo encoding into binary codewords and are subsequently transmitted to the fusion center via error-prone reporting channels. Upon exploiting the locally most powerful test (LMPT) strategy, we devise two multi-bit LMPT detectors in which quantized raw observations and local likelihood ratios are fused respectively. Moreover, the asymptotic detection performance of the proposed quantized detectors is analyzed, and closed-form expressions for the detection and false alarm probabilities are derived. Furthermore, the multi-bit quantizer design criterion, considering both RQ and LQ, is then proposed to achieve near-optimal asymptotic performance for our proposed detectors. The normalized Fisher information and asymptotic relative efficiency are derived, serving as tools to analyze and compensate for the loss of information introduced by the quantization. Simulation results validate the effectiveness of the proposed detectors, especially in scenarios with low signal-to-noise ratios and poor channel conditions.
Feeding power amplifiers (PAs) with constant envelope (CE) signals is an effective way to reduce the power consumption in massive multiple-input-multiple-output (MIMO) systems. The nonlinear distortion caused by CE signaling must be mitigated by means of signal processing to improve the achievable sum rates. To this purpose, many linear and nonlinear precoding techniques have been developed for the CE MIMO downlink. The vast majority of these CE precoding techniques do not include a power allocation scheme, which is indispensable to achieve adequate performances in the downlink with channel gain imbalances between users. In this paper, we present two algorithms to produce a power allocation scheme for regularized zero-forcing (RZF) precoding in CE MIMO downlink. Both techniques are based on transforming the CE quantized MIMO downlink to an approximately equivalent system of parallel single-input-single-output (SISO) channels. The first technique is proven to solve the sum rate maximization problem in the approximate system optimally, whereas the second technique obtains the local maximum with lower complexity. We also extend another state-of-the-art quantization aware sum rate maximization algorithm with linear precoding to the CE downlink. Numerical results illustrate significant gains for the performance of the RZF precoder when the CE quantization is taken into account in a power allocation. Another key numerical result is that the proposed RZF techniques achieve almost the identical performance so that the one with lower computational complexity is chosen as the main method. Results also show that the proposed RZF precoding schemes perform at least as good as the state-of-the-art method with an advantage that the main RZF method has significantly lower computational complexity than the state-of-the-art.
While recent advancements have highlighted the role of low-resolution analog-to-digital converters (ADCs) in integrated sensing and communication (ISAC) systems, the specific impact of ADC resolution on hybrid radar fusion (HRF), where we fuse monostatic and bistatic sensing systems remains relatively unexplored. The uplink (UL) paths in HRF, comprising both direct and reflected signals within the same frequency band, pose unique challenges, particularly given that the reflected signal is often significantly weaker than the direct path, making HRF systems susceptible to ADC resolution. To investigate the influence of quantization and ADC resolution on HRF, we employ the quantized Cram\'er-Rao bound (CRB) as a metric for sensing accuracy. This work derives the quantized CRB specifically for HRF systems and the quantized communication rate. We extend our analysis to obtain lower bounds on the Fisher Information Matrix (FIM) and UL communication rate, which we use to characterize quantized HRF systems. Using these derived bounds, we analyze the quantized HRF system through the lens of CRB-rate boundaries. We obtain the CRB-rate boundary through two optimization problems, where each solution point represents a trade-off boundary between the sensing accuracy and the communication rate. Extensive simulations illustrate the influence of ADC resolution, dynamic range (DR), and various system parameters on the CRB-rate boundary of HRF systems. These results offer critical insights into the design of efficient and high-performance HRF systems.