We consider a decentralized learning problem, where a set of computing nodes aim at solving a non-convex optimization problem collaboratively. It is well-known that decentralized optimization schemes face two major system bottlenecks: stragglers' delay and communication overhead. In this paper, we tackle these bottlenecks by proposing a novel decentralized and gradient-based optimization algorithm named as QuanTimed-DSGD. Our algorithm stands on two main ideas: (i) we impose a deadline on the local gradient computations of each node at each iteration of the algorithm, and (ii) the nodes exchange quantized versions of their local models. The first idea robustifies to straggling nodes and the second alleviates communication efficiency. The key technical contribution of our work is to prove that with non-vanishing noises for quantization and stochastic gradients, the proposed method exactly converges to the global optimal for convex loss functions, and finds a first-order stationary point in non-convex scenarios. Our numerical evaluations of the QuanTimed-DSGD on training benchmark datasets, MNIST and CIFAR-10, demonstrate speedups of up to 3x in run-time, compared to state-of-the-art decentralized optimization methods.
A major challenge facing Computer Vision systems is providing the ability to accurately detect threats and recognize subjects and/or objects under dynamically changing network conditions. We propose two novel models that characterize the face recognition accuracy in terms of video encoding parameters. Specifically, we model the accuracy in terms of video resolution, quantization, and actual bit rate. We validate the models using two distinct video datasets and a large image dataset by conducting 1, 668 experiments that involve simultaneously varying combinations of encoding parameters. We show that both models hold true for the deep learning and statistical based face recognition. Furthermore, we show that the models can be used to capture different accuracy metrics, specifically the recall, precision, and F1-score. Ultimately, we provide meaningful insights on the factors affecting the constants of each proposed model.
Typical convolutional networks are trained and conducted on RGB images. However, images are often compressed for memory savings and efficient transmission in real-world applications. In this paper, we explore methods for performing semantic segmentation on the discrete cosine transform (DCT) representation defined by the JPEG standard. We first rearrange the DCT coefficients to form a preferred input type, then we tailor an existing network to the DCT inputs. The proposed method has an accuracy close to the RGB model at about the same network complexity. Moreover, we investigate the impact of selecting different DCT components on segmentation performance. With a proper selection, one can achieve the same level accuracy using only 36% of the DCT coefficients. We further show the robustness of our method under the quantization errors. To our knowledge, this paper is the first to explore semantic segmentation on the DCT representation.
Due to their high computational complexity, deep neural networks are still limited to powerful processing units. To promote a reduced model complexity by dint of low-bit fixed-point quantization, we propose a gradient-based optimization strategy to generate a symmetric mixture of Gaussian modes (SGM) where each mode belongs to a particular quantization stage. We achieve 2-bit state-of-the-art performance and illustrate the model's ability for self-dependent weight adaptation during training.
Model compression has become necessary when applying neural networks (NN) into many real application tasks that can accept slightly-reduced model accuracy with strict tolerance to model complexity. Recently, Knowledge Distillation, which distills the knowledge from well-trained and highly complex teacher model into a compact student model, has been widely used for model compression. However, under the strict requirement on the resource cost, it is quite challenging to achieve comparable performance with the teacher model, essentially due to the drastically-reduced expressiveness ability of the compact student model. Inspired by the nature of the expressiveness ability in Neural Networks, we propose to use multi-segment activation, which can significantly improve the expressiveness ability with very little cost, in the compact student model. Specifically, we propose a highly efficient multi-segment activation, called Light Multi-segment Activation (LMA), which can rapidly produce multiple linear regions with very few parameters by leveraging the statistical information. With using LMA, the compact student model is capable of achieving much better performance effectively and efficiently, than the ReLU-equipped one with same model scale. Furthermore, the proposed method is compatible with other model compression techniques, such as quantization, which means they can be used jointly for better compression performance. Experiments on state-of-the-art NN architectures over the real-world tasks demonstrate the effectiveness and extensibility of the LMA.
Due to a resource-constrained environment, network compression has become an important part of deep neural networks research. In this paper, we propose a new compression method, \textit{Inter-Layer Weight Prediction} (ILWP) and quantization method which quantize the predicted residuals between the weights in all convolution layers based on an inter-frame prediction method in conventional video coding schemes. Furthermore, we found a phenomenon \textit{Smoothly Varying Weight Hypothesis} (SVWH) which is that the weights in adjacent convolution layers share strong similarity in shapes and values, i.e., the weights tend to vary smoothly along with the layers. Based on SVWH, we propose a second ILWP and quantization method which quantize the predicted residuals between the weights in adjacent convolution layers. Since the predicted weight residuals tend to follow Laplace distributions with very low variance, the weight quantization can more effectively be applied, thus producing more zero weights and enhancing the weight compression ratio. In addition, we propose a new \textit{inter-layer loss} for eliminating non-texture bits, which enabled us to more effectively store only texture bits. That is, the proposed loss regularizes the weights such that the collocated weights between the adjacent two layers have the same values. Finally, we propose an ILWP with an inter-layer loss and quantization method. Our comprehensive experiments show that the proposed method achieves a much higher weight compression rate at the same accuracy level compared with the previous quantization-based compression methods in deep neural networks.
To make music composition more approachable, we designed the first AI-powered Google Doodle, the Bach Doodle, where users can create their own melody and have it harmonized by a machine learning model Coconet (Huang et al., 2017) in the style of Bach. For users to input melodies, we designed a simplified sheet-music based interface. To support an interactive experience at scale, we re-implemented Coconet in TensorFlow.js (Smilkov et al., 2019) to run in the browser and reduced its runtime from 40s to 2s by adopting dilated depth-wise separable convolutions and fusing operations. We also reduced the model download size to approximately 400KB through post-training weight quantization. We calibrated a speed test based on partial model evaluation time to determine if the harmonization request should be performed locally or sent to remote TPU servers. In three days, people spent 350 years worth of time playing with the Bach Doodle, and Coconet received more than 55 million queries. Users could choose to rate their compositions and contribute them to a public dataset, which we are releasing with this paper. We hope that the community finds this dataset useful for applications ranging from ethnomusicological studies, to music education, to improving machine learning models.
In this paper, we address the problem of reducing the memory footprint of convolutional network architectures. We introduce a vector quantization method that aims at preserving the quality of the reconstruction of the network outputs rather than its weights. The principle of our approach is that it minimizes the loss reconstruction error for in-domain inputs. Our method only requires a set of unlabelled data at quantization time and allows for efficient inference on CPU by using byte-aligned codebooks to store the compressed weights. We validate our approach by quantizing a high performing ResNet-50 model to a memory size of 5MB (20x compression factor) while preserving a top-1 accuracy of 76.1% on ImageNet object classification and by compressing a Mask R-CNN with a 26x factor.
1 bit deep neural networks (DNNs), of which both the activations and weights are binarized , are attracting more and more attention due to their high computational efficiency and low memory requirement . However, the drawback of large accuracy dropping also restrict s its application. In this paper, we propose a novel Targeted Acceleration and Compression (TAC) framework to improve the performance of 1 bit deep neural networks W e consider that the acceleration and compression effects of binarizing fully connected layer s are not sufficient to compensate for the accuracy loss caused by it In the proposed framework, t he convolutional and fully connected layer are separated and optimized i ndividually . F or the convolutional layer s , both the activations and weights are binarized. For the fully connected layer s, the binarization operation is re placed by network pruning and low bit quantization. The proposed framework is implemented on the CIFAR 10, CIFAR 100 and ImageNet ( ILSVRC 12 ) datasets , and experimental results show that the proposed TAC can significantly improve the accuracy of 1 bit deep neural networks and outperforms the state of the art by more than 6 percentage points .
We propose and study a multi-scale approach to vector quantization. We develop an algorithm, dubbed reconstruction trees, inspired by decision trees. Here the objective is parsimonious reconstruction of unsupervised data, rather than classification. Contrasted to more standard vector quantization methods, such as K-means, the proposed approach leverages a family of given partitions, to quickly explore the data in a coarse to fine-- multi-scale-- fashion. Our main technical contribution is an analysis of the expected distortion achieved by the proposed algorithm, when the data are assumed to be sampled from a fixed unknown distribution. In this context, we derive both asymptotic and finite sample results under suitable regularity assumptions on the distribution. As a special case, we consider the setting where the data generating distribution is supported on a compact Riemannian sub-manifold. Tools from differential geometry and concentration of measure are useful in our analysis.
Coded caching provides significant gains over conventional uncoded caching by creating multicasting opportunities among distinct requests. Massive multiple-input multiple-output (MIMO) systems require downlink channel state information (CSI) at the base station (BS) to better utilize the available spatial diversity and multiplexing gains. However, in a frequency division duplex (FDD) massive MIMO system, the huge CSI feedback overhead becomes restrictive and degrades the overall spectral efficiency. In this paper, we propose a deep learning based channel state matrix compression scheme, called DeepCMC, composed of convolutional layers followed by quantization and entropy coding blocks. In comparison with previous works, the main contributions of DeepCMC are two-fold: i) DeepCMC is fully convolutional, and it can be used in a wide range of scenarios with various numbers of sub-channels and transmit antennas; ii) DeepCMC includes quantization and entropy coding blocks and minimizes a cost function that accounts for both the rate of compression and the reconstruction quality of the channel matrix at the BS. Simulation results demonstrate that DeepCMC significantly outperforms the state of the art compression schemes in terms of the reconstruction quality of the channel state matrix for the same compression rate, measured in bits per channel dimension.
Large deep neural network (DNN) models pose the key challenge to energy efficiency due to the significantly higher energy consumption of off-chip DRAM accesses than arithmetic or SRAM operations. It motivates the intensive research on model compression with two main approaches. Weight pruning leverages the redundancy in the number of weights and can be performed in a non-structured, which has higher flexibility and pruning rate but incurs index accesses due to irregular weights, or structured manner, which preserves the full matrix structure with lower pruning rate. Weight quantization leverages the redundancy in the number of bits in weights. Compared to pruning, quantization is much more hardware-friendly, and has become a "must-do" step for FPGA and ASIC implementations. This paper provides a definitive answer to the question for the first time. First, we build ADMM-NN-S by extending and enhancing ADMM-NN, a recently proposed joint weight pruning and quantization framework. Second, we develop a methodology for fair and fundamental comparison of non-structured and structured pruning in terms of both storage and computation efficiency. Our results show that ADMM-NN-S consistently outperforms the prior art: (i) it achieves 348x, 36x, and 8x overall weight pruning on LeNet-5, AlexNet, and ResNet-50, respectively, with (almost) zero accuracy loss; (ii) we demonstrate the first fully binarized (for all layers) DNNs can be lossless in accuracy in many cases. These results provide a strong baseline and credibility of our study. Based on the proposed comparison framework, with the same accuracy and quantization, the results show that non-structrued pruning is not competitive in terms of both storage and computation efficiency. Thus, we conclude that non-structured pruning is considered harmful. We urge the community not to continue the DNN inference acceleration for non-structured sparsity.
In this paper we tackle the problem of recovering the phase of complex linear measurements when only magnitude information is available and we control the input. We are motivated by the recent development of dedicated optics-based hardware for rapid random projections which leverages the propagation of light in random media. A signal of interest $\mathbf{\xi} \in \mathbb{R}^N$ is mixed by a random scattering medium to compute the projection $\mathbf{y} = \mathbf{A} \mathbf{\xi}$, with $\mathbf{A} \in \mathbb{C}^{M \times N}$ being a realization of a standard complex Gaussian iid random matrix. Such optics-based matrix multiplications can be much faster and energy-efficient than their CPU or GPU counterparts, yet two difficulties must be resolved: only the intensity ${|\mathbf{y}|}^2$ can be recorded by the camera, and the transmission matrix $\mathbf{A}$ is unknown. We show that even without knowing $\mathbf{A}$, we can recover the unknown phase of $\mathbf{y}$ for some equivalent transmission matrix with the same distribution as $\mathbf{A}$. Our method is based on two observations: first, conjugating or changing the phase of any row of $\mathbf{A}$ does not change its distribution; and second, since we control the input we can interfere $\mathbf{\xi}$ with arbitrary reference signals. We show how to leverage these observations to cast the measurement phase retrieval problem as a Euclidean distance geometry problem. We demonstrate appealing properties of the proposed algorithm in both numerical simulations and real hardware experiments. Not only does our algorithm accurately recover the missing phase, but it mitigates the effects of quantization and the sensitivity threshold, thus improving the measured magnitudes.
Acoustic Event Detection (AED), aiming at detecting categories of events based on audio signals, has found application in many intelligent systems. Recently deep neural network significantly advances this field and reduces detection errors to a large scale. However how to efficiently execute deep models in AED has received much less attention. Meanwhile state-of-the-art AED models are based on large deep models, which are computational demanding and challenging to deploy on devices with constrained computational resources. In this paper, we present a simple yet effective compression approach which jointly leverages knowledge distillation and quantization to compress larger network (teacher model) into compact network (student model). Experimental results show proposed technique not only lowers error rate of original compact network by 15% through distillation but also further reduces its model size to a large extent (2% of teacher, 12% of full-precision student) through quantization.
With the development of deep neural networks, the size of network models becomes larger and larger. Model compression has become an urgent need for deploying these network models to mobile or embedded devices. Model quantization is a representative model compression technique. Although a lot of quantization methods have been proposed, many of them suffer from a high quantization error caused by a long-tail distribution of network weights. In this paper, we propose a novel quantization method, called weight normalization based quantization (WNQ), for model compression. WNQ adopts weight normalization to avoid the long-tail distribution of network weights and subsequently reduces the quantization error. Experiments on CIFAR-100 and ImageNet show that WNQ can outperform other baselines to achieve state-of-the-art performance.
Visual localization algorithms have achieved significant improvements in performance thanks to recent advances in camera technology and vision-based techniques. However, there remains one critical caveat: all current approaches that are based on image retrieval currently scale at best linearly with the size of the environment with respect to both storage, and consequentially in most approaches, query time. This limitation severely curtails the capability of autonomous systems in a wide range of compute, power, storage, size, weight or cost constrained applications such as drones. In this work, we present a novel binary tree encoding approach for visual localization which can serve as an alternative for existing quantization and indexing techniques. The proposed tree structure allows us to derive a compressed training scheme that achieves sub-linearity in both required storage and inference time. The encoding memory can be easily configured to satisfy different storage constraints. Moreover, our approach is amenable to an optional sequence filtering mechanism to further improve the localization results, while maintaining the same amount of storage. Our system is entirely agnostic to the front-end descriptors, allowing it to be used on top of recent state-of-the-art image representations. Experimental results show that the proposed method significantly outperforms state-of-the-art approaches under limited storage constraints.
Radiologists use time series of medical images to monitor the progression of a patient condition. They compare information gleaned from sequences of images to gain insight on progression or remission of the lesions, thus evaluating the progress of a patient condition or response to therapy. Visual methods of determining differences between one series of images to another can be subjective or fail to detect very small differences. We propose the use of quantization errors obtained from Self Organizing Maps for image content analysis. We tested this technique with MRI images to which we progressively added synthetic lesions. We have used a global approach that considers changes on the entire image as opposed to changes in segmented lesion regions only. We claim that this approach does not suffer from the limitations imposed by segmentation, which may compromise the results. Results show quantization errors increased with the increase in lesions on the images. The results are also consistent with previous studies using alternative approaches. We then compared the detectability ability of our method to that of human novice observers having to detect very small local differences in random-dot images. The quantization errors of the SOM outputs compared with correct positive rates, after subtraction of false positive rates (guess rates), increased noticeably and consistently with small increases in local dot size that were not detectable by humans. We conclude that our method detects very small changes in complex images and suggest that it could be implemented to assist human operators in image based decision making.
Binary Neural Networks (BNNs) show promising progress in reducing computational and memory costs but suffer from substantial accuracy degradation compared to their real-valued counterparts on large-scale datasets, e.g., ImageNet. Previous work mainly focused on reducing quantization errors of weights and activations, whereby a series of approximation methods and sophisticated training tricks have been proposed. In this work, we make several observations that challenge conventional wisdom. We revisit some commonly used techniques, such as scaling factors and custom gradients, and show that these methods are not crucial in training well-performing BNNs. On the contrary, we suggest several design principles for BNNs based on the insights learned and demonstrate that highly accurate BNNs can be trained from scratch with a simple training strategy. We propose a new BNN architecture BinaryDenseNet, which significantly surpasses all existing 1-bit CNNs on ImageNet without tricks. In our experiments, BinaryDenseNet achieves 18.6% and 7.6% relative improvement over the well-known XNOR-Network and the current state-of-the-art Bi-Real Net in terms of top-1 accuracy on ImageNet, respectively.
In this work, a deep learning-based quantization scheme for log-likelihood ratio (L-value) storage is introduced. We analyze the dependency between the average magnitude of different L-values from the same quadrature amplitude modulation (QAM) symbol and show they follow a consistent ordering. Based on this we design a deep autoencoder that jointly compresses and separately reconstructs each L-value, allowing the use of a weighted loss function that aims to more accurately reconstructs low magnitude inputs. Our method is shown to be competitive with state-of-the-art maximum mutual information quantization schemes, reducing the required memory footprint by a ratio of up to two and a loss of performance smaller than 0.1 dB with less than two effective bits per L-value or smaller than 0.04 dB with 2.25 effective bits. We experimentally show that our proposed method is a universal compression scheme in the sense that after training on an LDPC-coded Rayleigh fading scenario we can reuse the same network without further training on other channel models and codes while preserving the same performance benefits.
Quantization has been an effective technology in ANN (approximate nearest neighbour) search due to its high accuracy and fast search speed. To meet the requirement of different applications, there is always a trade-off between retrieval accuracy and speed, reflected by variable code lengths. However, to encode the dataset into different code lengths, existing methods need to train several models, where each model can only produce a specific code length. This incurs a considerable training time cost, and largely reduces the flexibility of quantization methods to be deployed in real applications. To address this issue, we propose a Deep Recurrent Quantization (DRQ) architecture which can generate sequential binary codes. To the end, when the model is trained, a sequence of binary codes can be generated and the code length can be easily controlled by adjusting the number of recurrent iterations. A shared codebook and a scalar factor is designed to be the learnable weights in the deep recurrent quantization block, and the whole framework can be trained in an end-to-end manner. As far as we know, this is the first quantization method that can be trained once and generate sequential binary codes. Experimental results on the benchmark datasets show that our model achieves comparable or even better performance compared with the state-of-the-art for image retrieval. But it requires significantly less number of parameters and training times. Our code is published online: https://github.com/cfm-uestc/DRQ.