Standard deep learning algorithms are implemented using floating-point real numbers. This presents an obstacle for implementing them on low-end devices which may not have dedicated floating-point units (FPUs). As a result, researchers in tinyML have considered machine learning algorithms that can train and run a deep neural network (DNN) on a low-end device using integer operations only. In this paper we propose PocketNN, a light and self-contained proof-of-concept framework in pure C++ for the training and inference of DNNs using only integers. Unlike other approaches, PocketNN directly operates on integers without requiring any explicit quantization algorithms or customized fixed-point formats. This was made possible by pocket activations, which are a family of activation functions devised for integer-only DNNs, and an emerging DNN training algorithm called direct feedback alignment (DFA). Unlike the standard backpropagation (BP), DFA trains each layer independently, thus avoiding integer overflow which is a key problem when using BP with integer-only operations. We used PocketNN to train some DNNs on two well-known datasets, MNIST and Fashion-MNIST. Our experiments show that the DNNs trained with our PocketNN achieved 96.98% and 87.7% accuracies on MNIST and Fashion-MNIST datasets, respectively. The accuracies are very close to the equivalent DNNs trained using BP with floating-point real number operations, such that accuracy degradations were just 1.02%p and 2.09%p, respectively. Finally, our PocketNN has high compatibility and portability for low-end devices as it is open source and implemented in pure C++ without any dependencies.
Analyzing numerous or long time series is difficult in practice due to the high storage costs and computational requirements. Therefore, techniques have been proposed to generate compact similarity-preserving representations of time series, enabling real-time similarity search on large in-memory data collections. However, the existing techniques are not ideally suited for assessing similarity when sequences are locally out of phase. In this paper, we propose the use of product quantization for efficient similarity-based comparison of time series under time warping. The idea is to first compress the data by partitioning the time series into equal length sub-sequences which are represented by a short code. The distance between two time series can then be efficiently approximated by pre-computed elastic distances between their codes. The partitioning into sub-sequences forces unwanted alignments, which we address with a pre-alignment step using the maximal overlap discrete wavelet transform (MODWT). To demonstrate the efficiency and accuracy of our method, we perform an extensive experimental evaluation on benchmark datasets in nearest neighbors classification and clustering applications. Overall, the proposed solution emerges as a highly efficient (both in terms of memory usage and computation time) replacement for elastic measures in time series applications.
Compact stellar systems such as Ultra-compact dwarfs (UCDs) and Globular Clusters (GCs) around galaxies are known to be the tracers of the merger events that have been forming these galaxies. Therefore, identifying such systems allows to study galaxies mass assembly, formation and evolution. However, in the lack of spectroscopic information detecting UCDs/GCs using imaging data is very uncertain. Here, we aim to train a machine learning model to separate these objects from the foreground stars and background galaxies using the multi-wavelength imaging data of the Fornax galaxy cluster in 6 filters, namely u, g, r, i, J and Ks. The classes of objects are highly imbalanced which is problematic for many automatic classification techniques. Hence, we employ Synthetic Minority Over-sampling to handle the imbalance of the training data. Then, we compare two classifiers, namely Localized Generalized Matrix Learning Vector Quantization (LGMLVQ) and Random Forest (RF). Both methods are able to identify UCDs/GCs with a precision and a recall of >93 percent and provide relevances that reflect the importance of each feature dimension %(colors and angular sizes) for the classification. Both methods detect angular sizes as important markers for this classification problem. While it is astronomical expectation that color indices of u-i and i-Ks are the most important colors, our analysis shows that colors such as g-r are more informative, potentially because of higher signal-to-noise ratio. Besides the excellent performance the LGMLVQ method allows further interpretability by providing the feature importance for each individual class, class-wise representative samples and the possibility for non-linear visualization of the data as demonstrated in this contribution. We conclude that employing machine learning techniques to identify UCDs/GCs can lead to promising results.
Conventional video compression (VC) methods are based on motion compensated transform coding, and the steps of motion estimation, mode and quantization parameter selection, and entropy coding are optimized individually due to the combinatorial nature of the end-to-end optimization problem. Learned VC allows end-to-end rate-distortion (R-D) optimized training of nonlinear transform, motion and entropy model simultaneously. Most works on learned VC consider end-to-end optimization of a sequential video codec based on R-D loss averaged over pairs of successive frames. It is well-known in conventional VC that hierarchical, bi-directional coding outperforms sequential compression because of its ability to use both past and future reference frames. This paper proposes a learned hierarchical bi-directional video codec (LHBDC) that combines the benefits of hierarchical motion-compensated prediction and end-to-end optimization. Experimental results show that we achieve the best R-D results that are reported for learned VC schemes to date in both PSNR and MS-SSIM. Compared to conventional video codecs, the R-D performance of our end-to-end optimized codec outperforms those of both x265 and SVT-HEVC encoders ("veryslow" preset) in PSNR and MS-SSIM as well as HM 16.23 reference software in MS-SSIM. We present ablation studies showing performance gains due to proposed novel tools such as learned masking, flow-field subsampling, and temporal flow vector prediction. The models and instructions to reproduce our results can be found in https://github.com/makinyilmaz/LHBDC/
Neural video codecs have demonstrated great potential in video transmission and storage applications. Existing neural hybrid video coding approaches rely on optical flow or Gaussian-scale flow for prediction, which cannot support fine-grained adaptation to diverse motion content. Towards more content-adaptive prediction, we propose a novel cross-scale prediction module that achieves more effective motion compensation. Specifically, on the one hand, we produce a reference feature pyramid as prediction sources and then transmit cross-scale flows that leverage the feature scale to control the precision of prediction. On the other hand, for the first time, a weighted prediction mechanism is introduced even if only a single reference frame is available, which can help synthesize a fine prediction result by transmitting cross-scale weight maps. In addition to the cross-scale prediction module, we further propose a multi-stage quantization strategy, which improves the rate-distortion performance with no extra computational penalty during inference. We show the encouraging performance of our efficient neural video codec (ENVC) on several benchmark datasets. In particular, the proposed ENVC can compete with the latest coding standard H.266/VVC in terms of sRGB PSNR on UVG dataset for the low-latency mode. We also analyze in detail the effectiveness of the cross-scale prediction module in handling various video content, and provide a comprehensive ablation study to analyze those important components. Test code is available at https://github.com/USTC-IMCL/ENVC .
This work considers a challenging Deep Neural Network(DNN) quantization task that seeks to train quantized DNNs without involving any full-precision operations. Most previous quantization approaches are not applicable to this task since they rely on full-precision gradients to update network weights. To fill this gap, in this work we advocate using Evolutionary Algorithms (EAs) to search for the optimal low-bits weights of DNNs. To efficiently solve the induced large-scale discrete problem, we propose a novel EA based on cooperative coevolution that repeatedly groups the network weights based on the confidence in their values and focuses on optimizing the ones with the least confidence. To the best of our knowledge, this is the first work that applies EAs to train quantized DNNs. Experiments show that our approach surpasses previous quantization approaches and can train a 4-bit ResNet-20 on the Cifar-10 dataset with the same test accuracy as its full-precision counterpart.
Quantized neural networks typically require smaller memory footprints and lower computation complexity, which is crucial for efficient deployment. However, quantization inevitably leads to a distribution divergence from the original network, which generally degrades the performance. To tackle this issue, massive efforts have been made, but most existing approaches lack statistical considerations and depend on several manual configurations. In this paper, we present an adaptive-mapping quantization method to learn an optimal latent sub-distribution that is inherent within models and smoothly approximated with a concrete Gaussian Mixture (GM). In particular, the network weights are projected in compliance with the GM-approximated sub-distribution. This sub-distribution evolves along with the weight update in a co-tuning schema guided by the direct task-objective optimization. Sufficient experiments on image classification and object detection over various modern architectures demonstrate the effectiveness, generalization property, and transferability of the proposed method. Besides, an efficient deployment flow for the mobile CPU is developed, achieving up to 7.46$\times$ inference acceleration on an octa-core ARM CPU. Our codes have been publicly released at \url{https://github.com/RunpeiDong/DGMS}.
Running large-scale pre-trained language models in computationally constrained environments remains a challenging problem yet to be addressed, while transfer learning from these models has become prevalent in Natural Language Processing tasks. Several solutions, including knowledge distillation, network quantization, or network pruning have been previously proposed; however, these approaches focus mostly on the English language, thus widening the gap when considering low-resource languages. In this work, we introduce three light and fast versions of distilled BERT models for the Romanian language: Distil-BERT-base-ro, Distil-RoBERT-base, and DistilMulti-BERT-base-ro. The first two models resulted from the individual distillation of knowledge from two base versions of Romanian BERTs available in literature, while the last one was obtained by distilling their ensemble. To our knowledge, this is the first attempt to create publicly available Romanian distilled BERT models, which were thoroughly evaluated on five tasks: part-of-speech tagging, named entity recognition, sentiment analysis, semantic textual similarity, and dialect identification. Our experimental results argue that the three distilled models offer performance comparable to their teachers, while being twice as fast on a GPU and ~35% smaller. In addition, we further test the similarity between the predictions of our students versus their teachers by measuring their label and probability loyalty, together with regression loyalty - a new metric introduced in this work.
In hash-based image retrieval systems, degraded or transformed inputs usually generate different codes from the original, deteriorating the retrieval accuracy. To mitigate this issue, data augmentation can be applied during training. However, even if augmented samples of an image are similar in real feature space, the quantization can scatter them far away in Hamming space. This results in representation discrepancies that can impede training and degrade performance. In this work, we propose a novel self-distilled hashing scheme to minimize the discrepancy while exploiting the potential of augmented data. By transferring the hash knowledge of the weakly-transformed samples to the strong ones, we make the hash code insensitive to various transformations. We also introduce hash proxy-based similarity learning and binary cross entropy-based quantization loss to provide fine quality hash codes. Ultimately, we construct a deep hashing framework that not only improves the existing deep hashing approaches, but also achieves the state-of-the-art retrieval results. Extensive experiments are conducted and confirm the effectiveness of our work.
Accelerating the neural network inference by FPGA has emerged as a popular option, since the reconfigurability and high performance computing capability of FPGA intrinsically satisfies the computation demand of the fast-evolving neural algorithms. However, the popular neural accelerators on FPGA (e.g., Xilinx DPU) mainly utilize the DSP resources for constructing their processing units, while the rich LUT resources are not well exploited. Via the software-hardware co-design approach, in this work, we develop an FPGA-based heterogeneous computing system for neural network acceleration. From the hardware perspective, the proposed accelerator consists of DSP- and LUT-based GEneral Matrix-Multiplication (GEMM) computing cores, which forms the entire computing system in a heterogeneous fashion. The DSP- and LUT-based GEMM cores are computed w.r.t a unified Instruction Set Architecture (ISA) and unified buffers. Along the data flow of the neural network inference path, the computation of the convolution/fully-connected layer is split into two portions, handled by the DSP- and LUT-based GEMM cores asynchronously. From the software perspective, we mathematically and systematically model the latency and resource utilization of the proposed heterogeneous accelerator, regarding varying system design configurations. Through leveraging the reinforcement learning technique, we construct a framework to achieve end-to-end selection and optimization of the design specification of target heterogeneous accelerator, including workload split strategy, mixed-precision quantization scheme, and resource allocation of DSP- and LUT-core. In virtue of the proposed design framework and heterogeneous computing system, our design outperforms the state-of-the-art Mix&Match design with latency reduced by 1.12-1.32x with higher inference accuracy. The N3H-core is open-sourced at: https://github.com/elliothe/N3H_Core.
Binary networks are extremely efficient as they use only two symbols to define the network: $\{+1,-1\}$. One can make the prior distribution of these symbols a design choice. The recent IR-Net of Qin et al. argues that imposing a Bernoulli distribution with equal priors (equal bit ratios) over the binary weights leads to maximum entropy and thus minimizes information loss. However, prior work cannot precisely control the binary weight distribution during training, and therefore cannot guarantee maximum entropy. Here, we show that quantizing using optimal transport can guarantee any bit ratio, including equal ratios. We investigate experimentally that equal bit ratios are indeed preferable and show that our method leads to optimization benefits. We show that our quantization method is effective when compared to state-of-the-art binarization methods, even when using binary weight pruning.
Deep hashing has shown promising performance in large-scale image retrieval. However, latent codes extracted by Deep Neural Networks (DNNs) will inevitably lose semantic information during the binarization process, which damages the retrieval accuracy and makes it challenging. Although many existing approaches perform regularization to alleviate quantization errors, we figure out an incompatible conflict between metric learning and quantization learning. The metric loss penalizes the inter-class distances to push different classes unconstrained far away. Worse still, it tends to map the latent code deviate from ideal binarization point and generate severe ambiguity in the binarization process. Based on the minimum distance of the binary linear code, we creatively propose Hashing-guided Hinge Function (HHF) to avoid such conflict. In detail, the carefully-designed inflection point, which relies on the hash bit length and category numbers, is explicitly adopted to balance the metric term and quantization term. Such a modification prevents the network from falling into local metric optimal minima in deep hashing. Extensive experiments in CIFAR-10, CIFAR-100, ImageNet, and MS-COCO show that HHF consistently outperforms existing techniques, and is robust and flexible to transplant into other methods. Code is available at https://github.com/JerryXu0129/HHF.
This is an exploratory study that discovers the current image quantization (vector quantization) do not satisfy translation equivariance in the quantized space due to aliasing. Instead of focusing on anti-aliasing, we propose a simple yet effective way to achieve translation-equivariant image quantization by enforcing orthogonality among the codebook embeddings. To explore the advantages of translation-equivariant image quantization, we conduct three proof-of-concept experiments with a carefully controlled dataset: (1) text-to-image generation, where the quantized image indices are the target to predict, (2) image-to-text generation, where the quantized image indices are given as a condition, (3) using a smaller training set to analyze sample efficiency. From the strictly controlled experiments, we empirically verify that the translation-equivariant image quantizer improves not only sample efficiency but also the accuracy over VQGAN up to +11.9% in text-to-image generation and +3.9% in image-to-text generation.
The nonuniform quantization strategy for compressing neural networks usually achieves better performance than its counterpart, i.e., uniform strategy, due to its superior representational capacity. However, many nonuniform quantization methods overlook the complicated projection process in implementing the nonuniformly quantized weights/activations, which incurs non-negligible time and space overhead in hardware deployment. In this study, we propose Nonuniform-to-Uniform Quantization (N2UQ), a method that can maintain the strong representation ability of nonuniform methods while being hardware-friendly and efficient as the uniform quantization for model inference. We achieve this through learning the flexible in-equidistant input thresholds to better fit the underlying distribution while quantizing these real-valued inputs into equidistant output levels. To train the quantized network with learnable input thresholds, we introduce a generalized straight-through estimator (G-STE) for intractable backward derivative calculation w.r.t. threshold parameters. Additionally, we consider entropy preserving regularization to further reduce information loss in weight quantization. Even under this adverse constraint of imposing uniformly quantized weights and activations, our N2UQ outperforms state-of-the-art nonuniform quantization methods by 0.5~1.7 on ImageNet, demonstrating the contribution of N2UQ design. Code and models are available at: https://github.com/liuzechun/Nonuniform-to-Uniform-Quantization.
Network quantization significantly reduces model inference complexity and has been widely used in real-world deployments. However, most existing quantization methods have been developed mainly on Convolutional Neural Networks (CNNs), and suffer severe degradation when applied to fully quantized vision transformers. In this work, we demonstrate that many of these difficulties arise because of serious inter-channel variation in LayerNorm inputs, and present, Power-of-Two Factor (PTF), a systematic method to reduce the performance degradation and inference complexity of fully quantized vision transformers. In addition, observing an extreme non-uniform distribution in attention maps, we propose Log-Int-Softmax (LIS) to sustain that and simplify inference by using 4-bit quantization and the BitShift operator. Comprehensive experiments on various transformer-based architectures and benchmarks show that our Fully Quantized Vision Transformer (FQ-ViT) outperforms previous works while even using lower bit-width on attention maps. For instance, we reach 84.89% top-1 accuracy with ViT-L on ImageNet and 50.8 mAP with Cascade Mask R-CNN (Swin-S) on COCO. To our knowledge, we are the first to achieve lossless accuracy degradation (~1%) on fully quantized vision transformers. The code is available at https://github.com/megvii-research/FQ-ViT.
Non-intrusive speech intelligibility (SI) prediction from binaural signals is useful in many applications. However, most existing signal-based measures are designed to be applied to single-channel signals. Measures specifically designed to take into account the binaural properties of the signal are often intrusive - characterised by requiring access to a clean speech signal - and typically rely on combining both channels into a single-channel signal before making predictions. This paper proposes a non-intrusive SI measure that computes features from a binaural input signal using a combination of vector quantization (VQ) and contrastive predictive coding (CPC) methods. VQ-CPC feature extraction does not rely on any model of the auditory system and is instead trained to maximise the mutual information between the input signal and output features. The computed VQ-CPC features are input to a predicting function parameterized by a neural network. Two predicting functions are considered in this paper. Both feature extractor and predicting functions are trained on simulated binaural signals with isotropic noise. They are tested on simulated signals with isotropic and real noise. For all signals, the ground truth scores are the (intrusive) deterministic binaural STOI. Results are presented in terms of correlations and MSE and demonstrate that VQ-CPC features are able to capture information relevant to modelling SI and outperform all the considered benchmarks - even when evaluating on data comprising of different noise field types.
Quantization is one of the most effective methods to compress neural networks, which has achieved great success on convolutional neural networks (CNNs). Recently, vision transformers have demonstrated great potential in computer vision. However, previous post-training quantization methods performed not well on vision transformer, resulting in more than 1% accuracy drop even in 8-bit quantization. Therefore, we analyze the problems of quantization on vision transformers. We observe the distributions of activation values after softmax and GELU functions are quite different from the Gaussian distribution. We also observe that common quantization metrics, such as MSE and cosine distance, are inaccurate to determine the optimal scaling factor. In this paper, we propose the twin uniform quantization method to reduce the quantization error on these activation values. And we propose to use a Hessian guided metric to evaluate different scaling factors, which improves the accuracy of calibration at a small cost. To enable the fast quantization of vision transformers, we develop an efficient framework, PTQ4ViT. Experiments show the quantized vision transformers achieve near-lossless prediction accuracy (less than 0.5% drop at 8-bit quantization) on the ImageNet classification task.
Network quantization is a dominant paradigm of model compression. However, the abrupt changes in quantized weights during training often lead to severe loss fluctuations and result in a sharp loss landscape, making the gradients unstable and thus degrading the performance. Recently, Sharpness-Aware Minimization (SAM) has been proposed to smooth the loss landscape and improve the generalization performance of the models. Nevertheless, directly applying SAM to the quantized models can lead to perturbation mismatch or diminishment issues, resulting in suboptimal performance. In this paper, we propose a novel method, dubbed Sharpness-Aware Quantization (SAQ), to explore the effect of SAM in model compression, particularly quantization for the first time. Specifically, we first provide a unified view of quantization and SAM by treating them as introducing quantization noises and adversarial perturbations to the model weights, respectively. According to whether the noise and perturbation terms depend on each other, SAQ can be formulated into three cases, which are analyzed and compared comprehensively. Furthermore, by introducing an efficient training strategy, SAQ only incurs a little additional training overhead compared with the default optimizer (e.g., SGD or AdamW). Extensive experiments on both convolutional neural networks and Transformers across various datasets (i.e., ImageNet, CIFAR-10/100, Oxford Flowers-102, Oxford-IIIT Pets) show that SAQ improves the generalization performance of the quantized models, yielding the SOTA results in uniform quantization. For example, on ImageNet, SAQ outperforms AdamW by 1.2% on the Top-1 accuracy for 4-bit ViT-B/16. Our 4-bit ResNet-50 surpasses the previous SOTA method by 0.9% on the Top-1 accuracy.
With the recent demand of deploying neural network models on mobile and edge devices, it is desired to improve the model's generalizability on unseen testing data, as well as enhance the model's robustness under fixed-point quantization for efficient deployment. Minimizing the training loss, however, provides few guarantees on the generalization and quantization performance. In this work, we fulfill the need of improving generalization and quantization performance simultaneously by theoretically unifying them under the framework of improving the model's robustness against bounded weight perturbation and minimizing the eigenvalues of the Hessian matrix with respect to model weights. We therefore propose HERO, a Hessian-enhanced robust optimization method, to minimize the Hessian eigenvalues through a gradient-based training process, simultaneously improving the generalization and quantization performance. HERO enables up to a 3.8% gain on test accuracy, up to 30% higher accuracy under 80% training label perturbation, and the best post-training quantization accuracy across a wide range of precision, including a >10% accuracy improvement over SGD-trained models for common model architectures on various datasets.
There has been an explosion of interest in designing high-performance Transformers. While Transformers have delivered significant performance improvements, training such networks is extremely memory intensive owing to storing all intermediate activations that are needed for gradient computation during backpropagation, especially for long sequences. To this end, we present Mesa, a memory-saving training framework for Transformers. Specifically, Mesa uses exact activations during forward pass while storing a low-precision version of activations to reduce memory consumption during training. The low-precision activations are then dequantized during back-propagation to compute gradients. Besides, to address the heterogeneous activation distributions in the multi-head self-attention layers, we propose a head-wise activation quantization strategy, which quantizes activations based on the statistics of each head to minimize the approximation error. To further boost training efficiency, we learn quantization parameters by running estimates. More importantly, by re-investing the saved memory in employing a larger batch size or scaling up model size, we may further improve the performance under constrained computational resources. Extensive experiments on ImageNet, CIFAR-100 and ADE20K demonstrate that Mesa can achieve flexible memory-savings (up to 50%) during training while achieving comparable or even better performance. Code is available at https://github.com/ziplab/Mesa.