Deep hashing methods have shown great retrieval accuracy and efficiency in large-scale image retrieval. How to optimize discrete hash bits is always the focus in deep hashing methods. A common strategy in these methods is to adopt an activation function, e.g. $\operatorname{sigmoid}(\cdot)$ or $\operatorname{tanh}(\cdot)$, and minimize a quantization loss to approximate discrete values. However, this paradigm may make more and more hash bits stuck into the wrong saturated area of the activation functions and never escaped. We call this problem "Dead Bits Problem~(DBP)". Besides, the existing quantization loss will aggravate DBP as well. In this paper, we propose a simple but effective gradient amplifier which acts before activation functions to alleviate DBP. Moreover, we devise an error-aware quantization loss to further alleviate DBP. It avoids the negative effect of quantization loss based on the similarity between two images. The proposed gradient amplifier and error-aware quantization loss are compatible with a variety of deep hashing methods. Experimental results on three datasets demonstrate the efficiency of the proposed gradient amplifier and the error-aware quantization loss.
Banding artifacts are artificially-introduced contours arising from the quantization of a smooth region in a video. Despite the advent of recent higher quality video systems with more efficient codecs, these artifacts remain conspicuous, especially on larger displays. In this work, a comprehensive subjective study is performed to understand the dependence of the banding visibility on encoding parameters and dithering. We subsequently develop a simple and intuitive no-reference banding index called CAMBI (Contrast-aware Multiscale Banding Index) which uses insights from Contrast Sensitivity Function in the Human Visual System to predict banding visibility. CAMBI correlates well with subjective perception of banding while using only a few visually-motivated hyperparameters.
In this paper, we develop a new classification method for manifold-valued data in the framework of probabilistic learning vector quantization. In many classification scenarios, the data can be naturally represented by symmetric positive definite matrices, which are inherently points that live on a curved Riemannian manifold. Due to the non-Euclidean geometry of Riemannian manifolds, traditional Euclidean machine learning algorithms yield poor results on such data. In this paper, we generalize the probabilistic learning vector quantization algorithm for data points living on the manifold of symmetric positive definite matrices equipped with Riemannian natural metric (affine-invariant metric). By exploiting the induced Riemannian distance, we derive the probabilistic learning Riemannian space quantization algorithm, obtaining the learning rule through Riemannian gradient descent. Empirical investigations on synthetic data, image data , and motor imagery EEG data demonstrate the superior performance of the proposed method.
Machine Learning compilers like TVM allow a fast and flexible deployment on embedded CPUs. This enables the use of non-standard operators, which are common in ML compression techniques. However, it is necessary to understand the limitations of typical compute-intense operators in ML workloads to design a proper solution. This is the first in-detail analysis of dense and convolution operators, generated with TVM, that compares to the fundamental hardware limits of embedded ARM processors. Thereby it explains the gap between computational peak performance, theoretical and measured, and real-world state-of-the-art results, created with TVM and openBLAS. Instead, one can see that single-precision general matrix multiply (GEMM) and convolutions are bound by L1-cache-read bandwidth. Explorations of 8-bit and bit-serial quantized operators show that quantization can be used to achieve relevant speedups compared to cache-bound floating-point operators. However, the performance of quantized operators highly depends on the interaction between data layout and bit packing.
The Posit Number System (PNS) was introduced by John L. Gustafson in 2017. The interesting properties of this novel format can be exploited under the scenario of deep neural networks. In this paper, we propose Deep PeNSieve, a framework for entirely performing both training and inference of deep neural networks employing the PNS. Furthermore, an 8-bit posit quantization approach using fused operations is introduced. In comparison with the state-of-the-art posit frameworks, the proposal has been able to train more complex networks than the feedforward ones, achieving similar accuracies as the floating-point format. The case of CIFAR-10 is especially remarkable, as 16-bit posits even obtain 4% higher top-1 for such dataset. Overall, results show that the proposed quantization approach can preserve model accuracy in the same manner as common quantization techniques.
Deep neural networks have been applied in many applications exhibiting extraordinary abilities in the field of computer vision. However, complex network architectures challenge efficient real-time deployment and require significant computation resources and energy costs. These challenges can be overcome through optimizations such as network compression. Network compression can often be realized with little loss of accuracy. In some cases accuracy may even improve. This paper provides a survey on two types of network compression: pruning and quantization. Pruning can be categorized as static if it is performed offline or dynamic if it is performed at run-time. We compare pruning techniques and describe criteria used to remove redundant computations. We discuss trade-offs in element-wise, channel-wise, shape-wise, filter-wise, layer-wise and even network-wise pruning. Quantization reduces computations by reducing the precision of the datatype. Weights, biases, and activations may be quantized typically to 8-bit integers although lower bit width implementations are also discussed including binary neural networks. Both pruning and quantization can be used independently or combined. We compare current techniques, analyze their strengths and weaknesses, present compressed network accuracy results on a number of frameworks, and provide practical guidance for compressing networks.
Adversarial examples contain carefully crafted perturbations that can fool deep neural networks (DNNs) into making wrong predictions. Enhancing the adversarial robustness of DNNs has gained considerable interest in recent years. Although image transformation-based defenses were widely considered at an earlier time, most of them have been defeated by adaptive attacks. In this paper, we propose a new image transformation defense based on error diffusion halftoning, and combine it with adversarial training to defend against adversarial examples. Error diffusion halftoning projects an image into a 1-bit space and diffuses quantization error to neighboring pixels. This process can remove adversarial perturbations from a given image while maintaining acceptable image quality in the meantime in favor of recognition. Experimental results demonstrate that the proposed method is able to improve adversarial robustness even under advanced adaptive attacks, while most of the other image transformation-based defenses do not. We show that a proper image transformation can still be an effective defense approach. Code: https://github.com/shaoyuanlo/Halftoning-Defense
Convolutional neural networks (CNN) have been widely used for boosting the performance of many machine intelligence tasks. However, the CNN models are usually computationally intensive and energy consuming, since they are often designed with numerous multiply-operations and considerable parameters for the accuracy reason. Thus, it is difficult to directly apply them in the resource-constrained environments such as 'Internet of Things' (IoT) devices and smart phones. To reduce the computational complexity and energy burden, here we present a novel minimalist hardware architecture using adder convolutional neural network (AdderNet), in which the original convolution is replaced by adder kernel using only additions. To maximally excavate the potential energy consumption, we explore the low-bit quantization algorithm for AdderNet with shared-scaling-factor method, and we design both specific and general-purpose hardware accelerators for AdderNet. Experimental results show that the adder kernel with int8/int16 quantization also exhibits high performance, meanwhile consuming much less resources (theoretically ~81% off). In addition, we deploy the quantized AdderNet on FPGA (Field Programmable Gate Array) platform. The whole AdderNet can practically achieve 16% enhancement in speed, 67.6%-71.4% decrease in logic resource utilization and 47.85%-77.9% decrease in power consumption compared to CNN under the same circuit architecture. With a comprehensive comparison on the performance, power consumption, hardware resource consumption and network generalization capability, we conclude the AdderNet is able to surpass all the other competitors including the classical CNN, novel memristor-network, XNOR-Net and the shift-kernel based network, indicating its great potential in future high performance and energy-efficient artificial intelligence applications.
Learning generative models is challenging for a network edge node with limited data and computing power. Since tasks in similar environments share model similarity, it is plausible to leverage pre-trained generative models from the cloud or other edge nodes. Appealing to optimal transport theory tailored towards Wasserstein-1 generative adversarial networks (WGAN), this study aims to develop a framework which systematically optimizes continual learning of generative models using local data at the edge node while exploiting adaptive coalescence of pre-trained generative models. Specifically, by treating the knowledge transfer from other nodes as Wasserstein balls centered around their pre-trained models, continual learning of generative models is cast as a constrained optimization problem, which is further reduced to a Wasserstein-1 barycenter problem. A two-stage approach is devised accordingly: 1) The barycenters among the pre-trained models are computed offline, where displacement interpolation is used as the theoretic foundation for finding adaptive barycenters via a "recursive" WGAN configuration; 2) the barycenter computed offline is used as meta-model initialization for continual learning and then fast adaptation is carried out to find the generative model using the local samples at the target edge node. Finally, a weight ternarization method, based on joint optimization of weights and threshold for quantization, is developed to compress the generative model further.
Federated learning (FL) enables multiple clients to collaboratively train a shared model without disclosing their local datasets. This is achieved by exchanging local model updates with the help of a parameter server (PS). However, due to the increasing size of the trained models, the communication load due to the iterative exchanges between the clients and the PS often becomes a bottleneck in the performance. Sparse communication is often employed to reduce the communication load, where only a small subset of the model updates are communicated from the clients to the PS. In this paper, we introduce a novel time-correlated sparsification (TCS) scheme, which builds upon the notion that sparse communication framework can be considered as identifying the most significant elements of the underlying model. Hence, TCS seeks a certain correlation between the sparse representations used at consecutive iterations in FL, so that the overhead due to encoding and transmission of the sparse representation can be significantly reduced without compromising the test accuracy. Through extensive simulations on the CIFAR-10 dataset, we show that TCS can achieve centralized training accuracy with 100 times sparsification, and up to 2000 times reduction in the communication load when employed together with quantization.
Convolutional neural networks are able to learn realistic image priors from numerous training samples in low-level image generation and restoration. We show that, for high-level image recognition tasks, we can further reconstruct "realistic" images of each category by leveraging intrinsic Batch Normalization (BN) statistics without any training data. Inspired by the popular VAE/GAN methods, we regard the zero-shot optimization process of synthetic images as generative modeling to match the distribution of BN statistics. The generated images serve as a calibration set for the following zero-shot network quantizations. Our method meets the needs for quantizing models based on sensitive information, \textit{e.g.,} due to privacy concerns, no data is available. Extensive experiments on benchmark datasets show that, with the help of generated data, our approach consistently outperforms existing data-free quantization methods.
Neural data compression has been shown to outperform classical methods in terms of $RD$ performance, with results still improving rapidly. At a high level, neural compression is based on an autoencoder that tries to reconstruct the input instance from a (quantized) latent representation, coupled with a prior that is used to losslessly compress these latents. Due to limitations on model capacity and imperfect optimization and generalization, such models will suboptimally compress test data in general. However, one of the great strengths of learned compression is that if the test-time data distribution is known and relatively low-entropy (e.g. a camera watching a static scene, a dash cam in an autonomous car, etc.), the model can easily be finetuned or adapted to this distribution, leading to improved $RD$ performance. In this paper we take this concept to the extreme, adapting the full model to a single video, and sending model updates (quantized and compressed using a parameter-space prior) along with the latent representation. Unlike previous work, we finetune not only the encoder/latents but the entire model, and - during finetuning - take into account both the effect of model quantization and the additional costs incurred by sending the model updates. We evaluate an image compression model on I-frames (sampled at 2 fps) from videos of the Xiph dataset, and demonstrate that full-model adaptation improves $RD$ performance by ~1 dB, with respect to encoder-only finetuning.
The last few years have seen gigantic leaps in algorithms and systems to support efficient deep learning inference. Pruning and quantization algorithms can now consistently compress neural networks by an order of magnitude. For a compressed neural network, a multitude of inference frameworks have been designed to maximize the performance of the target hardware. While we find mature support for quantized neural networks in production frameworks such as OpenVINO and MNN, support for pruned sparse neural networks is still lacking. To tackle this challenge, we present SparseDNN, a sparse deep learning inference engine targeting CPUs. We present both kernel-level optimizations with a sparse code generator to accelerate sparse operators and novel network-level optimizations catering to sparse networks. We show that our sparse code generator can achieve significant speedups over state-of-the-art sparse and dense libraries. On end-to-end benchmarks such as Huggingface pruneBERT, SparseDNN achieves up to 5x throughput improvement over dense inference with state-of-the-art OpenVINO. Open source library at: https://github.com/marsupialtail/sparsednn.
Camera-based Deep Learning algorithms are increasingly needed for perception in Automated Driving systems. However, constraints from the automotive industry challenge the deployment of CNNs by imposing embedded systems with limited computational resources. In this paper, we propose an approach to embed a multi-task CNN network under such conditions on a commercial prototype platform, i.e. a low power System on Chip (SoC) processing four surround-view fisheye cameras at 10 FPS. The first focus is on designing an efficient and compact multi-task network architecture. Secondly, a pruning method is applied to compress the CNN, helping to reduce the runtime and memory usage by a factor of 2 without lowering the performances significantly. Finally, several embedded optimization techniques such as mixed-quantization format usage and efficient data transfers between different memory areas are proposed to ensure real-time execution and avoid bandwidth bottlenecks. The approach is evaluated on the hardware platform, considering embedded detection performances, runtime and memory bandwidth. Unlike most works from the literature that focus on classification task, we aim here to study the effect of pruning and quantization on a compact multi-task network with object detection, semantic segmentation and soiling detection tasks.
In this paper, we approach the problem of optimizing blackbox functions over large hybrid search spaces consisting of both combinatorial and continuous parameters. We demonstrate that previous evolutionary algorithms which rely on mutation-based approaches, while flexible over combinatorial spaces, suffer from a curse of dimensionality in high dimensional continuous spaces both theoretically and empirically, which thus limits their scope over hybrid search spaces as well. In order to combat this curse, we propose ES-ENAS, a simple and modular joint optimization procedure combining the class of sample-efficient smoothed gradient techniques, commonly known as Evolutionary Strategies (ES), with combinatorial optimizers in a highly scalable and intuitive way, inspired by the one-shot or supernet paradigm introduced in Efficient Neural Architecture Search (ENAS). By doing so, we achieve significantly more sample efficiency, which we empirically demonstrate over synthetic benchmarks, and are further able to apply ES-ENAS for architecture search over popular RL benchmarks.
Artificial neural network has achieved the state-of-art performance in fault detection on the Tennessee Eastman process, but it often requires enormous memory to fund its massive parameters. In order to implement online real-time fault detection, three deep compression techniques (pruning, clustering, and quantization) are applied to reduce the computational burden. We have extensively studied 7 different combinations of compression techniques, all methods achieve high model compression rates over 64% while maintain high fault detection accuracy. The best result is applying all three techniques, which reduces the model sizes by 91.5% and remains a high accuracy over 94%. This result leads to a smaller storage requirement in production environments, and makes the deployment smoother in real world.
Recently, transformer-based language models such as BERT have shown tremendous performance improvement for a range of natural language processing tasks. However, these language models usually are computation expensive and memory intensive during inference. As a result, it is difficult to deploy them on resource-restricted devices. To improve the inference performance, as well as reduce the model size while maintaining the model accuracy, we propose a novel quantization method named KDLSQ-BERT that combines knowledge distillation (KD) with learned step size quantization (LSQ) for language model quantization. The main idea of our method is that the KD technique is leveraged to transfer the knowledge from a "teacher" model to a "student" model when exploiting LSQ to quantize that "student" model during the quantization training process. Extensive experiment results on GLUE benchmark and SQuAD demonstrate that our proposed KDLSQ-BERT not only performs effectively when doing different bit (e.g. 2-bit $\sim$ 8-bit) quantization, but also outperforms the existing BERT quantization methods, and even achieves comparable performance as the full-precision base-line model while obtaining 14.9x compression ratio. Our code will be public available.
Deep learning models typically use single-precision (FP32) floating point data types for representing activations and weights, but a slew of recent research work has shown that computations with reduced-precision data types (FP16, 16-bit integers, 8-bit integers or even 4- or 2-bit integers) are enough to achieve same accuracy as FP32 and are much more efficient. Therefore, we designed fbgemm, a high-performance kernel library, from ground up to perform high-performance quantized inference on current generation CPUs. fbgemm achieves efficiency by fusing common quantization operations with a high-performance gemm implementation and by shape- and size-specific kernel code generation at runtime. The library has been deployed at Facebook, where it delivers greater than 2x performance gains with respect to our current production baseline.
Integer quantization of neural networks can be defined as the approximation of the high precision computation of the canonical neural network formulation, using reduced integer precision. It plays a significant role in the efficient deployment and execution of machine learning (ML) systems, reducing memory consumption and leveraging typically faster computations. In this work, we present an integer-only quantization strategy for Long Short-Term Memory (LSTM) neural network topologies, which themselves are the foundation of many production ML systems. Our quantization strategy is accurate (e.g. works well with quantization post-training), efficient and fast to execute (utilizing 8 bit integer weights and mostly 8 bit activations), and is able to target a variety of hardware (by leveraging instructions sets available in common CPU architectures, as well as available neural accelerators).
We introduce an automated tool for deploying ultra low-latency, low-power deep neural networks with convolutional layers on FPGAs. By extending the hls4ml library, we demonstrate an inference latency of $5\,\mu$s using convolutional architectures, targeting microsecond latency applications like those at the CERN Large Hadron Collider. Considering benchmark models trained on the Street View House Numbers Dataset, we demonstrate various methods for model compression in order to fit the computational constraints of a typical FPGA device used in trigger and data acquisition systems of particle detectors. In particular, we discuss pruning and quantization-aware training, and demonstrate how resource utilization can be significantly reduced with little to no loss in model accuracy. We show that the FPGA critical resource consumption can be reduced by 97% with zero loss in model accuracy, and by 99% when tolerating a 6% accuracy degradation.