Large Language Models~(LLMs) have become foundational in the realm of natural language processing, demonstrating performance improvements as model sizes increase. The Mixture-of-Experts~(MoE) approach offers a promising way to scale LLMs more efficiently by using fewer computational FLOPs through sparse activation. However, it suffers from significant memory overheads, necessitating model compression techniques. Post-training quantization, a popular method for model compression, proves less effective when directly applied to MoE models due to MoE's overlooked inherent sparsity. This paper explores several MoE structure-aware quantization heuristics, ranging from coarse to fine granularity, from MoE block to individual linear weight. Our investigations reveal critical principles: different MoE structures (i.e., blocks, experts, linear layers) require varying numbers of weight bits for effective and efficient quantization. Conclusions are supported by extensive benchmarking across two representative MoE models and six tasks. We further introduce novel enhancements to more accurately identify the most critical weights in MoE quantization that necessitate higher bit allocations, including the linear weight outlier scorer and MoE block scorer. Additionally, subsequent experiments validate our findings in the context of both weight and activation quantization.
Recent studies in joint source-channel coding (JSCC) have fostered a fresh paradigm in end-to-end semantic communication. Despite notable performance achievements, present initiatives in building semantic communication systems primarily hinge on the transmission of continuous channel symbols, thus presenting challenges in compatibility with established digital systems. In this paper, we introduce a novel approach to address this challenge by developing a multi-order digital joint coding-modulation (MDJCM) scheme for semantic communications. Initially, we construct a digital semantic communication system by integrating a multi-order modulation/demodulation module into a nonlinear transform source-channel coding (NTSCC) framework. Recognizing the non-differentiable nature of modulation/demodulation, we propose a novel substitution training strategy. Herein, we treat modulation/demodulation as a constrained quantization process and introduce scaling operations alongside manually crafted noise to approximate this process. As a result, employing this approximation in training semantic communication systems can be deployed in practical modulation/demodulation scenarios with superior performance. Additionally, we demonstrate the equivalence by analyzing the involved probability distribution. Moreover, to further upgrade the performance, we develop a hierarchical dimension-reduction strategy to provide a gradual information extraction process. Extensive experimental evaluations demonstrate the superiority of our proposed method over existing digital and non-digital JSCC techniques.
Historically, most speech models in machine-learning have used the mel-spectrogram as a speech representation. Recently, discrete audio tokens produced by neural audio codecs have become a popular alternate speech representation for speech synthesis tasks such as text-to-speech (TTS). However, the data distribution produced by such codecs is too complex for some TTS models to predict, typically requiring large autoregressive models to get good quality. Most existing audio codecs use Residual Vector Quantization (RVQ) to compress and reconstruct the time-domain audio signal. We propose a spectral codec which uses Finite Scalar Quantization (FSQ) to compress the mel-spectrogram and reconstruct the time-domain audio signal. A study of objective audio quality metrics and subjective listening tests suggests that our spectral codec has comparable perceptual quality to equivalent audio codecs. We show that FSQ, and the use of spectral speech representations, can both improve the performance of parallel TTS models.
In virtual reality (VR) applications, 360-degree images play a pivotal role in crafting immersive experiences and offering panoramic views, thus improving user Quality of Experience (QoE). However, the voluminous data generated by 360-degree images poses challenges in network storage and bandwidth. To address these challenges, we propose a novel Activation Map-based Vector Quantization (AM-VQ) framework, which is designed to reduce communication overhead for wireless transmission. The proposed AM-VQ scheme uses the Deep Neural Networks (DNNs) with vector quantization (VQ) to extract and compress semantic features. Particularly, the AM-VQ framework utilizes activation map to adaptively quantize semantic features, thus reducing data distortion caused by quantization operation. To further enhance the reconstruction quality of the 360-degree image, adversarial training with a Generative Adversarial Networks (GANs) discriminator is incorporated. Numerical results show that our proposed AM-VQ scheme achieves better performance than the existing Deep Learning (DL) based coding and the traditional coding schemes under the same transmission symbols.
Modern speaker verification (SV) systems typically demand expensive storage and computing resources, thereby hindering their deployment on mobile devices. In this paper, we explore adaptive neural network quantization for lightweight speaker verification. Firstly, we propose a novel adaptive uniform precision quantization method which enables the dynamic generation of quantization centroids customized for each network layer based on k-means clustering. By applying it to the pre-trained SV systems, we obtain a series of quantized variants with different bit widths. To enhance the performance of low-bit quantized models, a mixed precision quantization algorithm along with a multi-stage fine-tuning (MSFT) strategy is further introduced. Unlike uniform precision quantization, mixed precision approach allows for the assignment of varying bit widths to different network layers. When bit combination is determined, MSFT is employed to progressively quantize and fine-tune network in a specific order. Finally, we design two distinct binary quantization schemes to mitigate performance degradation of 1-bit quantized models: the static and adaptive quantizers. Experiments on VoxCeleb demonstrate that lossless 4-bit uniform precision quantization is achieved on both ResNets and DF-ResNets, yielding a promising compression ratio of around 8. Moreover, compared to uniform precision approach, mixed precision quantization not only obtains additional performance improvements with a similar model size but also offers the flexibility to generate bit combination for any desirable model size. In addition, our suggested 1-bit quantization schemes remarkably boost the performance of binarized models. Finally, a thorough comparison with existing lightweight SV systems reveals that our proposed models outperform all previous methods by a large margin across various model size ranges.
Large language models (LLMs) have achieved remarkable performance on Natural Language Processing (NLP) tasks, but they are hindered by high computational costs and memory requirements. Ternarization, an extreme form of quantization, offers a solution by reducing memory usage and enabling energy-efficient floating-point additions. However, applying ternarization to LLMs faces challenges stemming from outliers in both weights and activations. In this work, observing asymmetric outliers and non-zero means in weights, we introduce Dual Learnable Ternarization (DLT), which enables both scales and shifts to be learnable. We also propose Outlier-Friendly Feature Knowledge Distillation (OFF) to recover the information lost in extremely low-bit quantization. The proposed OFF can incorporate semantic information and is insensitive to outliers. At the core of OFF is maximizing the mutual information between features in ternarized and floating-point models using cosine similarity. Extensive experiments demonstrate that our TernaryLLM surpasses previous low-bit quantization methods on the standard text generation and zero-shot benchmarks for different LLM families. Specifically, for one of the most powerful open-source models, LLaMA-3, our approach (W1.58A16) outperforms the previous state-of-the-art method (W2A16) by 5.8 in terms of perplexity on C4 and by 8.2% in terms of average accuracy on zero-shot tasks.
We propose a new transformer-based image and video tokenizer with Binary Spherical Quantization (BSQ). BSQ projects the high-dimensional visual embedding to a lower-dimensional hypersphere and then applies binary quantization. BSQ is (1) parameter-efficient without an explicit codebook, (2) scalable to arbitrary token dimensions, and (3) compact: compressing visual data by up to 100$\times$ with minimal distortion. Our tokenizer uses a transformer encoder and decoder with simple block-wise causal masking to support variable-length videos as input. The resulting BSQ-ViT achieves state-of-the-art visual reconstruction quality on image and video reconstruction benchmarks with 2.4$\times$ throughput compared to the best prior methods. Furthermore, by learning an autoregressive prior for adaptive arithmetic coding, BSQ-ViT achieves comparable results on video compression with state-of-the-art video compression standards. BSQ-ViT also enables masked language models to achieve competitive image synthesis quality to GAN- and diffusion-based methods.
Low-bit quantization has become widespread for compressing image super-resolution (SR) models for edge deployment, which allows advanced SR models to enjoy compact low-bit parameters and efficient integer/bitwise constructions for storage compression and inference acceleration, respectively. However, it is notorious that low-bit quantization degrades the accuracy of SR models compared to their full-precision (FP) counterparts. Despite several efforts to alleviate the degradation, the transformer-based SR model still suffers severe degradation due to its distinctive activation distribution. In this work, we present a dual-stage low-bit post-training quantization (PTQ) method for image super-resolution, namely 2DQuant, which achieves efficient and accurate SR under low-bit quantization. The proposed method first investigates the weight and activation and finds that the distribution is characterized by coexisting symmetry and asymmetry, long tails. Specifically, we propose Distribution-Oriented Bound Initialization (DOBI), using different searching strategies to search a coarse bound for quantizers. To obtain refined quantizer parameters, we further propose Distillation Quantization Calibration (DQC), which employs a distillation approach to make the quantized model learn from its FP counterpart. Through extensive experiments on different bits and scaling factors, the performance of DOBI can reach the state-of-the-art (SOTA) while after stage two, our method surpasses existing PTQ in both metrics and visual effects. 2DQuant gains an increase in PSNR as high as 4.52dB on Set5 (x2) compared with SOTA when quantized to 2-bit and enjoys a 3.60x compression ratio and 5.08x speedup ratio. The code and models will be available at https://github.com/Kai-Liu001/2DQuant.
In this work, we propose a two-stage sign language production (SLP) paradigm that first encodes sign language sequences into discrete codes and then autoregressively generates sign language from text based on the learned codebook. However, existing vector quantization (VQ) methods are fixed-length encodings, overlooking the uneven information density in sign language, which leads to under-encoding of important regions and over-encoding of unimportant regions. To address this issue, we propose a novel dynamic vector quantization (DVA-VAE) model that can dynamically adjust the encoding length based on the information density in sign language to achieve accurate and compact encoding. Then, a GPT-like model learns to generate code sequences and their corresponding durations from spoken language text. Extensive experiments conducted on the PHOENIX14T dataset demonstrate the effectiveness of our proposed method. To promote sign language research, we propose a new large German sign language dataset, PHOENIX-News, which contains 486 hours of sign language videos, audio, and transcription texts.Experimental analysis on PHOENIX-News shows that the performance of our model can be further improved by increasing the size of the training data. Our project homepage is https://t2sgpt-demo.yinaoxiong.cn.
This paper introduces new methodology based on the field of Topological Data Analysis for detecting anomalies in multivariate time series, that aims to detect global changes in the dependency structure between channels. The proposed approach is lean enough to handle large scale datasets, and extensive numerical experiments back the intuition that it is more suitable for detecting global changes of correlation structures than existing methods. Some theoretical guarantees for quantization algorithms based on dependent time sequences are also provided.
This paper explores the combination of neural network quantization and entropy coding for memory footprint minimization. Edge deployment of quantized models is hampered by the harsh Pareto frontier of the accuracy-to-bitwidth tradeoff, causing dramatic accuracy loss below a certain bitwidth. This accuracy loss can be alleviated thanks to mixed precision quantization, allowing for more flexible bitwidth allocation. However, standard mixed precision benefits remain limited due to the 1-bit frontier, that forces each parameter to be encoded on at least 1 bit of data. This paper introduces an approach that combines mixed precision, zero-point quantization and entropy coding to push the compression boundary of Resnets beyond the 1-bit frontier with an accuracy drop below 1% on the ImageNet benchmark. From an implementation standpoint, a compact decoder architecture features reduced latency, thus allowing for inference-compatible decoding.
Humans can effortlessly draw new categories from a single exemplar, a feat that has long posed a challenge for generative models. However, this gap has started to close with recent advances in diffusion models. This one-shot drawing task requires powerful inductive biases that have not been systematically investigated. Here, we study how different inductive biases shape the latent space of Latent Diffusion Models (LDMs). Along with standard LDM regularizers (KL and vector quantization), we explore supervised regularizations (including classification and prototype-based representation) and contrastive inductive biases (using SimCLR and redundancy reduction objectives). We demonstrate that LDMs with redundancy reduction and prototype-based regularizations produce near-human-like drawings (regarding both samples' recognizability and originality) -- better mimicking human perception (as evaluated psychophysically). Overall, our results suggest that the gap between humans and machines in one-shot drawings is almost closed.
Large language models (LLMs) are omnipresent, however their practical deployment is challenging due to their ever increasing computational and memory demands. Quantization is one of the most effective ways to make them more compute and memory efficient. Quantization-aware training (QAT) methods, generally produce the best quantized performance, however it comes at the cost of potentially long training time and excessive memory usage, making it impractical when applying for LLMs. Inspired by parameter-efficient fine-tuning (PEFT) and low-rank adaptation (LoRA) literature, we propose LR-QAT -- a lightweight and memory-efficient QAT algorithm for LLMs. LR-QAT employs several components to save memory without sacrificing predictive performance: (a) low-rank auxiliary weights that are aware of the quantization grid; (b) a downcasting operator using fixed-point or double-packed integers and (c) checkpointing. Unlike most related work, our method (i) is inference-efficient, leading to no additional overhead compared to traditional PTQ; (ii) can be seen as a general extended pretraining framework, meaning that the resulting model can still be utilized for any downstream task afterwards; (iii) can be applied across a wide range of quantization settings, such as different choices quantization granularity, activation quantization, and seamlessly combined with many PTQ techniques. We apply LR-QAT to LLaMA-1/2/3 and Mistral model families and validate its effectiveness on several downstream tasks. Our method outperforms common post-training quantization (PTQ) approaches and reaches the same model performance as full-model QAT at the fraction of its memory usage. Specifically, we can train a 7B LLM on a single consumer grade GPU with 24GB of memory. Our source code is available at https://github.com/qualcomm-ai-research/LR-QAT
Winner-takes-all training is a simple learning paradigm, which handles ambiguous tasks by predicting a set of plausible hypotheses. Recently, a connection was established between Winner-takes-all training and centroidal Voronoi tessellations, showing that, once trained, hypotheses should quantize optimally the shape of the conditional distribution to predict. However, the best use of these hypotheses for uncertainty quantification is still an open question.In this work, we show how to leverage the appealing geometric properties of the Winner-takes-all learners for conditional density estimation, without modifying its original training scheme. We theoretically establish the advantages of our novel estimator both in terms of quantization and density estimation, and we demonstrate its competitiveness on synthetic and real-world datasets, including audio data.
The impressive capabilities in Large Language Models (LLMs) provide a powerful approach to reimagine users' typing experience. This paper demonstrates Proofread, a novel Gboard feature powered by a server-side LLM in Gboard, enabling seamless sentence-level and paragraph-level corrections with a single tap. We describe the complete system in this paper, from data generation, metrics design to model tuning and deployment. To obtain models with sufficient quality, we implement a careful data synthetic pipeline tailored to online use cases, design multifaceted metrics, employ a two-stage tuning approach to acquire the dedicated LLM for the feature: the Supervised Fine Tuning (SFT) for foundational quality, followed by the Reinforcement Learning (RL) tuning approach for targeted refinement. Specifically, we find sequential tuning on Rewrite and proofread tasks yields the best quality in SFT stage, and propose global and direct rewards in the RL tuning stage to seek further improvement. Extensive experiments on a human-labeled golden set showed our tuned PaLM2-XS model achieved 85.56\% good ratio. We launched the feature to Pixel 8 devices by serving the model on TPU v5 in Google Cloud, with thousands of daily active users. Serving latency was significantly reduced by quantization, bucket inference, text segmentation, and speculative decoding. Our demo could be seen in \href{https://youtu.be/4ZdcuiwFU7I}{Youtube}.
Diffusion-based image generation models have achieved great success in recent years by showing the capability of synthesizing high-quality content. However, these models contain a huge number of parameters, resulting in a significantly large model size. Saving and transferring them is a major bottleneck for various applications, especially those running on resource-constrained devices. In this work, we develop a novel weight quantization method that quantizes the UNet from Stable Diffusion v1.5 to 1.99 bits, achieving a model with 7.9X smaller size while exhibiting even better generation quality than the original one. Our approach includes several novel techniques, such as assigning optimal bits to each layer, initializing the quantized model for better performance, and improving the training strategy to dramatically reduce quantization error. Furthermore, we extensively evaluate our quantized model across various benchmark datasets and through human evaluation to demonstrate its superior generation quality.
Decentralized strategies are of interest for learning from large-scale data over networks. This paper studies learning over a network of geographically distributed nodes/agents subject to quantization. Each node possesses a private local cost function, collectively contributing to a global cost function, which the considered methodology aims to minimize. In contrast to many existing papers, the information exchange among nodes is log-quantized to address limited network-bandwidth in practical situations. We consider a first-order computationally efficient distributed optimization algorithm (with no extra inner consensus loop) that leverages node-level gradient correction based on local data and network-level gradient aggregation only over nearby nodes. This method only requires balanced networks with no need for stochastic weight design. It can handle log-scale quantized data exchange over possibly time-varying and switching network setups. We study convergence over both structured networks (for example, training over data-centers) and ad-hoc multi-agent networks (for example, training over dynamic robotic networks). Through experimental validation, we show that (i) structured networks generally result in a smaller optimality gap, and (ii) log-scale quantization leads to a smaller optimality gap compared to uniform quantization.
Large-scale universal speech models (USM) are already used in production. However, as the model size grows, the serving cost grows too. Serving cost of large models is dominated by model size that is why model size reduction is an important research topic. In this work we are focused on model size reduction using weights only quantization. We present the weights binarization of USM Recurrent Neural Network Transducer (RNN-T) and show that its model size can be reduced by 15.9x times at cost of word error rate (WER) increase by only 1.9% in comparison to the float32 model. It makes it attractive for practical applications.
Differential privacy mechanisms such as the Gaussian or Laplace mechanism have been widely used in data analytics for preserving individual privacy. However, they are mostly designed for continuous outputs and are unsuitable for scenarios where discrete values are necessary. Although various quantization mechanisms were proposed recently to generate discrete outputs under differential privacy, the outcomes are either biased or have an inferior accuracy-privacy trade-off. In this paper, we propose a family of quantization mechanisms that is unbiased and differentially private. It has a high degree of freedom and we show that some existing mechanisms can be considered as special cases of ours. To find the optimal mechanism, we formulate a linear optimization that can be solved efficiently using linear programming tools. Experiments show that our proposed mechanism can attain a better privacy-accuracy trade-off compared to baselines.
Deep Neural Networks (DNN) have become a promising paradigm when developing Artificial Intelligence (AI) and Machine Learning (ML) applications. However, DNN applications are vulnerable to fake data that are crafted with adversarial attack algorithms. Under adversarial attacks, the prediction accuracy of DNN applications suffers, making them unreliable. In order to defend against adversarial attacks, we introduce a novel noise-reduction procedure, Vector Quantization U-Net (VQUNet), to reduce adversarial noise and reconstruct data with high fidelity. VQUNet features a discrete latent representation learning through a multi-scale hierarchical structure for both noise reduction and data reconstruction. The empirical experiments show that the proposed VQUNet provides better robustness to the target DNN models, and it outperforms other state-of-the-art noise-reduction-based defense methods under various adversarial attacks for both Fashion-MNIST and CIFAR10 datasets. When there is no adversarial attack, the defense method has less than 1% accuracy degradation for both datasets.