Coresets are small, weighted summaries of larger datasets, aiming at providing provable error bounds for machine learning (ML) tasks while significantly reducing the communication and computation costs. To achieve a better trade-off between ML error bounds and costs, we propose the first framework to incorporate quantization techniques into the process of coreset construction. Specifically, we theoretically analyze the ML error bounds caused by a combination of coreset construction and quantization. Based on that, we formulate an optimization problem to minimize the ML error under a fixed budget of communication cost. To improve the scalability for large datasets, we identify two proxies of the original objective function, for which efficient algorithms are developed. For the case of data on multiple nodes, we further design a novel algorithm to allocate the communication budget to the nodes while minimizing the overall ML error. Through extensive experiments on multiple real-world datasets, we demonstrate the effectiveness and efficiency of our proposed algorithms for a variety of ML tasks. In particular, our algorithms have achieved more than 90% data reduction with less than 10% degradation in ML performance in most cases.
We propose a Composite Code Sparse Autoencoder (CCSA) approach for Approximate Nearest Neighbor (ANN) search of document representations based on Siamese-BERT models. In Information Retrieval (IR), the ranking pipeline is generally decomposed in two stages: the first stage focus on retrieving a candidate set from the whole collection. The second stage re-ranks the candidate set by relying on more complex models. Recently, Siamese-BERT models have been used as first stage ranker to replace or complement the traditional bag-of-word models. However, indexing and searching a large document collection require efficient similarity search on dense vectors and this is why ANN techniques come into play. Since composite codes are naturally sparse, we first show how CCSA can learn efficient parallel inverted index thanks to an uniformity regularizer. Second, CCSA can be used as a binary quantization method and we propose to combine it with the recent graph based ANN techniques. Our experiments on MSMARCO dataset reveal that CCSA outperforms IVF with product quantization. Furthermore, CCSA binary quantization is beneficial for the index size, and memory usage for the graph-based HNSW method, while maintaining a good level of recall and MRR. Third, we compare with recent supervised quantization methods for image retrieval and find that CCSA is able to outperform them.
Recent developments in communication technologies, such as 5G, together with innovative computing paradigms, such as edge computing, provide further possibilities for the implementation of real-time networked control systems. However, privacy and cyber-security concerns arise when sharing private data between sensors, agents and a third-party computing facility. In this paper, a secure version of the distributed formation control is presented, analyzed and simulated, where gradient-based formation control law is implemented in the edge, with sensor and actuator information being secured by fully homomorphic encryption method based on learning with error (FHE-LWE) combined with a proposed mixed uniform-logarithmic quantizer (MULQ). The novel quantizer is shown to be suitable for realizing secure control systems with FHE-LWE where the critical real-time information can be quantized into a prescribed bounded space of plaintext while satisfying a sector bound condition whose lower and upper-bound can be made sufficiently close to an identity. An absolute stability analysis is presented, that shows the asymptotic stability of the closed-loop secure control system.
Let $K: \boldsymbol{\Omega}\times \boldsymbol{\Omega}$ be a continuous Mercer kernel defined on a compact subset of ${\mathbb R}^n$ and $\mathcal{H}_K$ be the reproducing kernel Hilbert space (RKHS) associated with $K$. Given a finite measure $\nu$ on $\boldsymbol{\Omega}$, we investigate upper and lower bounds on the $\varepsilon$-entropy of the unit ball of $\mathcal{H}_K$ in the space $L_p(\nu)$. This topic is an important direction in the modern statistical theory of kernel-based methods. We prove sharp upper and lower bounds for $p\in [1,+\infty]$. For $p\in [1,2]$, the upper bounds are determined solely by the eigenvalue behaviour of the corresponding integral operator $\phi\to \int_{\boldsymbol{\Omega}} K(\cdot,{\mathbf y})\phi({\mathbf y})d\nu({\mathbf y})$. In constrast, for $p>2$, the bounds additionally depend on the convergence rate of the truncated Mercer series to the kernel $K$ in the $L_p(\nu)$-norm. We discuss a number of consequences of our bounds and show that they are substantially tighter than previous bounds for general kernels. Furthermore, for specific cases, such as zonal kernels and the Gaussian kernel on a box, our bounds are asymptotically tight as $\varepsilon\to +0$.
Deep Neural Networks (DNNs) have gained considerable attention in the past decades due to their astounding performance in different applications, such as natural language modeling, self-driving assistance, and source code understanding. With rapid exploration, more and more complex DNN architectures have been proposed along with huge pre-trained model parameters. The common way to use such DNN models in user-friendly devices (e.g., mobile phones) is to perform model compression before deployment. However, recent research has demonstrated that model compression, e.g., model quantization, yields accuracy degradation as well as outputs disagreements when tested on unseen data. Since the unseen data always include distribution shifts and often appear in the wild, the quality and reliability of quantized models are not ensured. In this paper, we conduct a comprehensive study to characterize and help users understand the behaviors of quantized models. Our study considers 4 datasets spanning from image to text, 8 DNN architectures including feed-forward neural networks and recurrent neural networks, and 42 shifted sets with both synthetic and natural distribution shifts. The results reveal that 1) data with distribution shifts happen more disagreements than without. 2) Quantization-aware training can produce more stable models than standard, adversarial, and Mixup training. 3) Disagreements often have closer top-1 and top-2 output probabilities, and $Margin$ is a better indicator than the other uncertainty metrics to distinguish disagreements. 4) Retraining with disagreements has limited efficiency in removing disagreements. We opensource our code and models as a new benchmark for further studying the quantized models.
Data-free quantization is a task that compresses the neural network to low bit-width without access to original training data. Most existing data-free quantization methods cause severe performance degradation due to inaccurate activation clipping range and quantization error, especially for low bit-width. In this paper, we present a simple yet effective data-free quantization method with accurate activation clipping and adaptive batch normalization. Accurate activation clipping (AAC) improves the model accuracy by exploiting accurate activation information from the full-precision model. Adaptive batch normalization firstly proposes to address the quantization error from distribution changes by updating the batch normalization layer adaptively. Extensive experiments demonstrate that the proposed data-free quantization method can yield surprisingly performance, achieving 64.33% top-1 accuracy of ResNet18 on ImageNet dataset, with 3.7% absolute improvement outperforming the existing state-of-the-art methods.
We propose an adaptive projection-gradient descent-shrinkage-splitting method (APGDSSM) to integrate penalty based channel pruning into quantization-aware training (QAT). APGDSSM concurrently searches weights in both the quantized subspace and the sparse subspace. APGDSSM uses shrinkage operator and a splitting technique to create sparse weights, as well as the Group Lasso penalty to push the weight sparsity into channel sparsity. In addition, we propose a novel complementary transformed l1 penalty to stabilize the training for extreme compression.
In intra coding, Rate Distortion Optimization (RDO) is performed to achieve the optimal intra mode from a pre-defined candidate list. The optimal intra mode is also required to be encoded and transmitted to the decoder side besides the residual signal, where lots of coding bits are consumed. To further improve the performance of intra coding in Versatile Video Coding (VVC), an intelligent intra mode derivation method is proposed in this paper, termed as Deep Learning based Intra Mode Derivation (DLIMD). In specific, the process of intra mode derivation is formulated as a multi-class classification task, which aims to skip the module of intra mode signaling for coding bits reduction. The architecture of DLIMD is developed to adapt to different quantization parameter settings and variable coding blocks including non-square ones, which are handled by one single trained model. Different from the existing deep learning based classification problems, the hand-crafted features are also fed into the intra mode derivation network besides the learned features from feature learning network. To compete with traditional method, one additional binary flag is utilized in the video codec to indicate the selected scheme with RDO. Extensive experimental results reveal that the proposed method can achieve 2.28%, 1.74%, and 2.18% bit rate reduction on average for Y, U, and V components on the platform of VVC test model, which outperforms the state-of-the-art works.
In this paper, we propose a novel prosody disentangle method for prosodic Text-to-Speech (TTS) model, which introduces the vector quantization (VQ) method to the auxiliary prosody encoder to obtain the decomposed prosody representations in an unsupervised manner. Rely on its advantages, the speaking styles, such as pitch, speaking velocity, local pitch variance, etc., are decomposed automatically into the latent quantize vectors. We also investigate the internal mechanism of VQ disentangle process by means of a latent variables counter and find that higher value dimensions usually represent prosody information. Experiments show that our model can control the speaking styles of synthesis results by directly manipulating the latent variables. The objective and subjective evaluations illustrated that our model outperforms the popular models.
Cancer is one of the deadliest diseases worldwide. Accurate diagnosis and classification of cancer subtypes are indispensable for effective clinical treatment. Promising results on automatic cancer subtyping systems have been published recently with the emergence of various deep learning methods. However, such automatic systems often overfit the data due to the high dimensionality and scarcity. In this paper, we propose to investigate automatic subtyping from an unsupervised learning perspective by directly constructing the underlying data distribution itself, hence sufficient data can be generated to alleviate the issue of overfitting. Specifically, we bypass the strong Gaussianity assumption that typically exists but fails in the unsupervised learning subtyping literature due to small-sized samples by vector quantization. Our proposed method better captures the latent space features and models the cancer subtype manifestation on a molecular basis, as demonstrated by the extensive experimental results.
Most AI projects start with a Python notebook running on a single laptop; however, one usually needs to go through a mountain of pains to scale it to handle larger dataset (for both experimentation and production deployment). These usually entail many manual and error-prone steps for the data scientists to fully take advantage of the available hardware resources (e.g., SIMD instructions, multi-processing, quantization, memory allocation optimization, data partitioning, distributed computing, etc.). To address this challenge, we have open sourced BigDL 2.0 at https://github.com/intel-analytics/BigDL/ under Apache 2.0 license (combining the original BigDL and Analytics Zoo projects); using BigDL 2.0, users can simply build conventional Python notebooks on their laptops (with possible AutoML support), which can then be transparently accelerated on a single node (with up-to 9.6x speedup in our experiments), and seamlessly scaled out to a large cluster (across several hundreds servers in real-world use cases). BigDL 2.0 has already been adopted by many real-world users (such as Mastercard, Burger King, Inspur, etc.) in production.
Binary Neural Networks (BNNs) are an extremely promising method to reduce deep neural networks' complexity and power consumption massively. Binarization techniques, however, suffer from ineligible performance degradation compared to their full-precision counterparts. Prior work mainly focused on strategies for sign function approximation during forward and backward phases to reduce the quantization error during the binarization process. In this work, we propose a Bi-Modal Distributed binarization method (\methodname{}). That imposes bi-modal distribution of the network weights by kurtosis regularization. The proposed method consists of a training scheme that we call Weight Distribution Mimicking (WDM), which efficiently imitates the full-precision network weight distribution to their binary counterpart. Preserving this distribution during binarization-aware training creates robust and informative binary feature maps and significantly reduces the generalization error of the BNN. Extensive evaluations on CIFAR-10 and ImageNet demonstrate the superiority of our method over current state-of-the-art schemes. Our source code, experimental settings, training logs, and binary models are available at \url{https://github.com/BlueAnon/BD-BNN}.
Large neural networks are difficult to deploy on mobile devices because of intensive computation and storage. To alleviate it, we study ternarization, a balance between efficiency and accuracy that quantizes both weights and activations into ternary values. In previous ternarized neural networks, a hard threshold {\Delta} is introduced to determine quantization intervals. Although the selection of {\Delta} greatly affects the training results, previous works estimate {\Delta} via an approximation or treat it as a hyper-parameter, which is suboptimal. In this paper, we present the Soft Threshold Ternary Networks (STTN), which enables the model to automatically determine quantization intervals instead of depending on a hard threshold. Concretely, we replace the original ternary kernel with the addition of two binary kernels at training time, where ternary values are determined by the combination of two corresponding binary values. At inference time, we add up the two binary kernels to obtain a single ternary kernel. Our method dramatically outperforms current state-of-the-arts, lowering the performance gap between full-precision networks and extreme low bit networks. Experiments on ImageNet with ResNet-18 (Top-1 66.2%) achieves new state-of-the-art. Update: In this version, we further fine-tune the experimental hyperparameters and training procedure. The latest STTN shows that ResNet-18 with ternary weights and ternary activations achieves up to 68.2% Top-1 accuracy on ImageNet. Code is available at: github.com/WeixiangXu/STTN.
Vector quantization (VQ) based ANN indexes, such as Inverted File System (IVF) and Product Quantization (PQ), have been widely applied to embedding based document retrieval thanks to the competitive time and memory efficiency. Originally, VQ is learned to minimize the reconstruction loss, i.e., the distortions between the original dense embeddings and the reconstructed embeddings after quantization. Unfortunately, such an objective is inconsistent with the goal of selecting ground-truth documents for the input query, which may cause severe loss of retrieval quality. Recent works identify such a defect, and propose to minimize the retrieval loss through contrastive learning. However, these methods intensively rely on queries with ground-truth documents, whose performance is limited by the insufficiency of labeled data. In this paper, we propose Distill-VQ, which unifies the learning of IVF and PQ within a knowledge distillation framework. In Distill-VQ, the dense embeddings are leveraged as "teachers", which predict the query's relevance to the sampled documents. The VQ modules are treated as the "students", which are learned to reproduce the predicted relevance, such that the reconstructed embeddings may fully preserve the retrieval result of the dense embeddings. By doing so, Distill-VQ is able to derive substantial training signals from the massive unlabeled data, which significantly contributes to the retrieval quality. We perform comprehensive explorations for the optimal conduct of knowledge distillation, which may provide useful insights for the learning of VQ based ANN index. We also experimentally show that the labeled data is no longer a necessity for high-quality vector quantization, which indicates Distill-VQ's strong applicability in practice.
Model quantization is considered as a promising method to greatly reduce the resource requirements of deep neural networks. To deal with the performance drop induced by quantization errors, a popular method is to use training data to fine-tune quantized networks. In real-world environments, however, such a method is frequently infeasible because training data is unavailable due to security, privacy, or confidentiality concerns. Zero-shot quantization addresses such problems, usually by taking information from the weights of a full-precision teacher network to compensate the performance drop of the quantized networks. In this paper, we first analyze the loss surface of state-of-the-art zero-shot quantization techniques and provide several findings. In contrast to usual knowledge distillation problems, zero-shot quantization often suffers from 1) the difficulty of optimizing multiple loss terms together, and 2) the poor generalization capability due to the use of synthetic samples. Furthermore, we observe that many weights fail to cross the rounding threshold during training the quantized networks even when it is necessary to do so for better performance. Based on the observations, we propose AIT, a simple yet powerful technique for zero-shot quantization, which addresses the aforementioned two problems in the following way: AIT i) uses a KL distance loss only without a cross-entropy loss, and ii) manipulates gradients to guarantee that a certain portion of weights are properly updated after crossing the rounding thresholds. Experiments show that AIT outperforms the performance of many existing methods by a great margin, taking over the overall state-of-the-art position in the field.
Dimension reduction and data quantization are two important methods for reducing data complexity. In the paper, we study the methodology of first reducing data dimension by random projection and then quantizing the projections to ternary or binary codes, which has been widely applied in classification. Usually, the quantization will seriously degrade the accuracy of classification due to high quantization errors. Interestingly, however, we observe that the quantization could provide comparable and often superior accuracy, as the data to be quantized are sparse features generated with common filters. Furthermore, this quantization property could be maintained in the random projections of sparse features, if both the features and random projection matrices are sufficiently sparse. By conducting extensive experiments, we validate and analyze this intriguing property.
Reducing the latency and model size has always been a significant research problem for live Automatic Speech Recognition (ASR) application scenarios. Along this direction, model quantization has become an increasingly popular approach to compress neural networks and reduce computation cost. Most of the existing practical ASR systems apply post-training 8-bit quantization. To achieve a higher compression rate without introducing additional performance regression, in this study, we propose to develop 4-bit ASR models with native quantization aware training, which leverages native integer operations to effectively optimize both training and inference. We conducted two experiments on state-of-the-art Conformer-based ASR models to evaluate our proposed quantization technique. First, we explored the impact of different precisions for both weight and activation quantization on the LibriSpeech dataset, and obtained a lossless 4-bit Conformer model with 5.8x size reduction compared to the float32 model. Following this, we for the first time investigated and revealed the viability of 4-bit quantization on a practical ASR system that is trained with large-scale datasets, and produced a lossless Conformer ASR model with mixed 4-bit and 8-bit weights that has 5x size reduction compared to the float32 model.
While several self-supervised approaches for learning discrete speech representation have been proposed, it is unclear how these seemingly similar approaches relate to each other. In this paper, we consider a generative model with discrete latent variables that learns a discrete representation for speech. The objective of learning the generative model is formulated as information-theoretic co-training. Besides the wide generality, the objective can be optimized with several approaches, subsuming HuBERT-like training and vector quantization for learning discrete representation. Empirically, we find that the proposed approach learns discrete representation that is highly correlated with phonetic units, more correlated than HuBERT-like training and vector quantization.
In this work, we consider a full-duplex (FD) massive multiple-input multiple-output (MIMO) cellular network with low-resolution analog-to-digital converters (ADCs) and digital-to-analog converter (DACs). Our first contribution is to provide a unified framework for reverse link (uplink) analysis where matched filters are applied at the FD base stations (BSs) under channel hardening. Second, we derive the expressions of the signal-to-quantization-plus-interference-plus-noise ratio (SQINR) for general and special cases. Finally, we quantify effects of quantization error, pilot contamination, and full duplexing for a hexagonal cell lattice on spectral efficiency and cumulative distribution function (CDF) to show that FD outperforms half duplex (HD) in a wide variety of scenarios.
Background: Skin cancer is one of the widely seen cancer worldwide and automatic classification of skin cancer can be benefited dermatology clinics for an accurate diagnosis. Hence, a machine learning-based automatic skin cancer detection model must be developed. Material and Method: This research interests to overcome automatic skin cancer detection problem. A colored skin cancer image dataset is used. This dataset contains 3297 images with two classes. An automatic multilevel textural and deep features-based model is presented. Multilevel fuse feature generation using discrete wavelet transform (DWT), local phase quantization (LPQ), local binary pattern (LBP), pre-trained DarkNet19, and DarkNet53 are utilized to generate features of the skin cancer images, top 1000 features are selected threshold value-based neighborhood component analysis (NCA). The chosen top 1000 features are classified using the 10-fold cross-validation technique. Results: To obtain results, ten-fold cross-validation is used and 91.54% classification accuracy results are obtained by using the recommended pyramidal hybrid feature generator and NCA selector-based model. Further, various training and testing separation ratios (90:10, 80:20, 70:30, 60:40, 50:50) are used and the maximum classification rate is calculated as 95.74% using the 90:10 separation ratio. Conclusions: The findings and accuracies calculated are denoted that this model can be used in dermatology and pathology clinics to simplify the skin cancer detection process and help physicians.