We consider a similarity measure between two sets $A$ and $B$ of vectors, that balances the average and maximum cosine distance between pairs of vectors, one from set $A$ and one from set $B$. As a motivation for this measure, we present lineage tracking in a database. To practically realize this measure, we need an approximate search algorithm that given a set of vectors $A$ and sets of vectors $B_1,...,B_n$, the algorithm quickly locates the set $B_i$ that maximizes the similarity measure. For the case where all sets are singleton sets, essentially each is a single vector, there are known efficient approximate search algorithms, e.g., approximated versions of tree search algorithms, locality-sensitive hashing (LSH), vector quantization (VQ) and proximity graph algorithms. In this work, we present approximate search algorithms for the general case. The underlying idea in these algorithms is encoding a set of vectors via a "long" single vector. The proposed approximate approach achieves significant performance gains over an optimized, exact search on vector sets.
Arguably the most common and salient object in daily video communications is the talking head, as encountered in social media, virtual classrooms, teleconferences, news broadcasting, talk shows, etc. When communication bandwidth is limited by network congestions or cost effectiveness, compression artifacts in talking head videos are inevitable. The resulting video quality degradation is highly visible and objectionable due to high acuity of human visual system to faces. To solve this problem, we develop a multi-modality deep convolutional neural network method for restoring face videos that are aggressively compressed. The main innovation is a new DCNN architecture that incorporates known priors of multiple modalities: the video-synchronized speech signal and semantic elements of the compression code stream, including motion vectors, code partition map and quantization parameters. These priors strongly correlate with the latent video and hence they are able to enhance the capability of deep learning to remove compression artifacts. Ample empirical evidences are presented to validate the superior performance of the proposed DCNN method on face videos over the existing state-of-the-art methods.
We formulate the entropy of a quantized artificial neural network as a differentiable function that can be plugged as a regularization term into the cost function minimized by gradient descent. Our formulation scales efficiently beyond the first order and is agnostic of the quantization scheme. The network can then be trained to minimize the entropy of the quantized parameters, so that they can be optimally compressed via entropy coding. We experiment with our entropy formulation at quantizing and compressing well-known network architectures over multiple datasets. Our approach compares favorably over similar methods, enjoying the benefits of higher order entropy estimate, showing flexibility towards non-uniform quantization (we use Lloyd-max quantization), scalability towards any entropy order to be minimized and efficiency in terms of compression. We show that HEMP is able to work in synergy with other approaches aiming at pruning or quantizing the model itself, delivering significant benefits in terms of storage size compressibility without harming the model's performance.
Model compression is generally performed by using quantization, low-rank approximation or pruning, for which various algorithms have been researched in recent years. One fundamental question is: what types of compression work better for a given model? Or even better: can we improve by combining compressions in a suitable way? We formulate this generally as a problem of optimizing the loss but where the weights are constrained to equal an additive combination of separately compressed parts; and we give an algorithm to learn the corresponding parts' parameters. Experimentally with deep neural nets, we observe that 1) we can find significantly better models in the error-compression space, indicating that different compression types have complementary benefits, and 2) the best type of combination depends exquisitely on the type of neural net. For example, we can compress ResNets and AlexNet using only 1 bit per weight without error degradation at the cost of adding a few floating point weights. However, VGG nets can be better compressed by combining low-rank with a few floating point weights.
Extracting and analyzing iris textures for biometric recognition has been extensively studied. As the transition of iris recognition from lab technology to nation-scale applications, most systems are facing high complexity in either time or space, leading to unfitness for embedded devices. In this paper, the proposed design includes a minimal set of computer vision modules and multi-mode QC-LDPC decoder which can alleviate variability and noise caused by iris acquisition and follow-up process. Several classes of QC-LDPC code from IEEE 802.16 are tested for the validity of accuracy improvement. Some of the codes mentioned above are used for further QC-LDPC decoder quantization, validation and comparison to each other. We show that we can apply Dynamic Partial Reconfiguration technology to implement the multi-mode QC-LDPC decoder for the iris recognition system. The results show that the implementation is power-efficient and good for edge applications.
With the development of image recovery models,especially those based on adversarial and perceptual losses,the detailed texture portions of images are being recovered more naturally.However,these restored images are similar but not identical in detail texture to their reference images.With traditional image quality assessment methods,results with better subjective perceived quality often score lower in objective scoring.Assessment methods suffer from subjective and objective inconsistencies.This paper proposes a regional differential information entropy (RDIE) method for image quality assessment to address this problem.This approach allows better assessment of similar but not identical textural details and achieves good agreement with perceived quality.Neural networks are used to reshape the process of calculating information entropy,improving the speed and efficiency of the operation. Experiments conducted with this study image quality assessment dataset and the PIPAL dataset show that the proposed RDIE method yields a high degree of agreement with people average opinion scores compared to other image quality assessment metrics,proving that RDIE can better quantify the perceived quality of images.
The 2D heatmap-based approaches have dominated Human Pose Estimation (HPE) for years due to high performance. However, the long-standing quantization error problem in the 2D heatmap-based methods leads to several well-known drawbacks: 1) The performance for the low-resolution inputs is limited; 2) To improve the feature map resolution for higher localization precision, multiple costly upsampling layers are required; 3) Extra post-processing is adopted to reduce the quantization error. To address these issues, we aim to explore a brand new scheme, called \textit{SimCC}, which reformulates HPE as two classification tasks for horizontal and vertical coordinates. The proposed SimCC uniformly divides each pixel into several bins, thus achieving \emph{sub-pixel} localization precision and low quantization error. Benefiting from that, SimCC can omit additional refinement post-processing and exclude upsampling layers under certain settings, resulting in a more simple and effective pipeline for HPE. Extensive experiments conducted over COCO, CrowdPose, and MPII datasets show that SimCC outperforms heatmap-based counterparts, especially in low-resolution settings by a large margin.
Cooperative beamforming across access points (APs) and fronthaul quantization strategies are essential for cloud radio access network (C-RAN) systems. The nonconvexity of the C-RAN optimization problems, which is stemmed from per-AP power and fronthaul capacity constraints, requires high computational complexity for executing iterative algorithms. To resolve this issue, we investigate a deep learning approach where the optimization module is replaced with a well-trained deep neural network (DNN). An efficient learning solution is proposed which constructs a DNN to produce a low-dimensional representation of optimal beamforming and quantization strategies. Numerical results validate the advantages of the proposed learning solution.
Deep learning has advanced from fully connected architectures to structured models organized into components, e.g., the transformer composed of positional elements, modular architectures divided into slots, and graph neural nets made up of nodes. In structured models, an interesting question is how to conduct dynamic and possibly sparse communication among the separate components. Here, we explore the hypothesis that restricting the transmitted information among components to discrete representations is a beneficial bottleneck. The motivating intuition is human language in which communication occurs through discrete symbols. Even though individuals have different understandings of what a "cat" is based on their specific experiences, the shared discrete token makes it possible for communication among individuals to be unimpeded by individual differences in internal representation. To discretize the values of concepts dynamically communicated among specialist components, we extend the quantization mechanism from the Vector-Quantized Variational Autoencoder to multi-headed discretization with shared codebooks and use it for discrete-valued neural communication (DVNC). Our experiments show that DVNC substantially improves systematic generalization in a variety of architectures -- transformers, modular architectures, and graph neural networks. We also show that the DVNC is robust to the choice of hyperparameters, making the method very useful in practice. Moreover, we establish a theoretical justification of our discretization process, proving that it has the ability to increase noise robustness and reduce the underlying dimensionality of the model.
We present the Power Law Graph Transformer, a transformer model with well defined deductive and inductive tasks for prediction and representation learning. The deductive task learns the dataset level (global) and instance level (local) graph structures in terms of learnable power law distribution parameters. The inductive task outputs the prediction probabilities using the deductive task output, similar to a transductive model. We trained our model with Turkish-English and Portuguese-English datasets from TED talk transcripts for machine translation and compared the model performance and characteristics to a transformer model with scaled dot product attention trained on the same experimental setup. We report BLEU scores of $17.79$ and $28.33$ on the Turkish-English and Portuguese-English translation tasks with our model, respectively. We also show how a duality between a quantization set and N-dimensional manifold representation can be leveraged to transform between local and global deductive-inductive outputs using successive application of linear and non-linear transformations end-to-end.
Recent years have witnessed a growing list of systems for distributed data-parallel training. Existing systems largely fit into two paradigms, i.e., parameter server and MPI-style collective operations. On the algorithmic side, researchers have proposed a wide range of techniques to lower the communication via system relaxations: quantization, decentralization, and communication delay. However, most, if not all, existing systems only rely on standard synchronous and asynchronous stochastic gradient (SG) based optimization, therefore, cannot take advantage of all possible optimizations that the machine learning community has been developing recently. Given this emerging gap between the current landscapes of systems and theory, we build BAGUA, a MPI-style communication library, providing a collection of primitives, that is both flexible and modular to support state-of-the-art system relaxation techniques of distributed training. Powered by this design, BAGUA has a great ability to implement and extend various state-of-the-art distributed learning algorithms. In a production cluster with up to 16 machines (128 GPUs), BAGUA can outperform PyTorch-DDP, Horovod and BytePS in the end-to-end training time by a significant margin (up to 2 times) across a diverse range of tasks. Moreover, we conduct a rigorous tradeoff exploration showing that different algorithms and system relaxations achieve the best performance over different network conditions.
We propose a novel neural model compression strategy combining data augmentation, knowledge transfer, pruning, and quantization for device-robust acoustic scene classification (ASC). Specifically, we tackle the ASC task in a low-resource environment leveraging a recently proposed advanced neural network pruning mechanism, namely Lottery Ticket Hypothesis (LTH), to find a sub-network neural model associated with a small amount non-zero model parameters. The effectiveness of LTH for low-complexity acoustic modeling is assessed by investigating various data augmentation and compression schemes, and we report an efficient joint framework for low-complexity multi-device ASC, called \emph{Acoustic Lottery}. Acoustic Lottery could compress an ASC model up to $1/10^{4}$ and attain a superior performance (validation accuracy of 79.4% and Log loss of 0.64) compared to its not compressed seed model. All results reported in this work are based on a joint effort of four groups, namely GT-USTC-UKE-Tencent, aiming to address the "Low-Complexity Acoustic Scene Classification (ASC) with Multiple Devices" in the DCASE 2021 Challenge Task 1a.
Quantization based model compression serves as high performing and fast approach for inference that yields models which are highly compressed when compared to their full-precision floating point counterparts. The most extreme quantization is a 1-bit representation of parameters such that they have only two possible values, typically -1(0) or +1, enabling efficient implementation of the ubiquitous dot product using only additions. The main contribution of this work is the introduction of a method to smooth the combinatorial problem of determining a binary vector of weights to minimize the expected loss for a given objective by means of empirical risk minimization with backpropagation. This is achieved by approximating a multivariate binary state over the weights utilizing a deterministic and differentiable transformation of real-valued, continuous parameters. The proposed method adds little overhead in training, can be readily applied without any substantial modifications to the original architecture, does not introduce additional saturating nonlinearities or auxiliary losses, and does not prohibit applying other methods for binarizing the activations. Contrary to common assertions made in the literature, it is demonstrated that binary weighted networks can train well with the same standard optimization techniques and similar hyperparameter settings as their full-precision counterparts, specifically momentum SGD with large learning rates and $L_2$ regularization. To conclude experiments demonstrate the method performs remarkably well across a number of inductive image classification tasks with various architectures compared to their full-precision counterparts. The source code is publicly available at https://bitbucket.org/YanivShu/binary_weighted_networks_public.
A prominent technique for reducing the memory footprint of Spiking Neural Networks (SNNs) without decreasing the accuracy significantly is quantization. However, the state-of-the-art only focus on employing the weight quantization directly from a specific quantization scheme, i.e., either the post-training quantization (PTQ) or the in-training quantization (ITQ), and do not consider (1) quantizing other SNN parameters (e.g., neuron membrane potential), (2) exploring different combinations of quantization approaches (i.e., quantization schemes, precision levels, and rounding schemes), and (3) selecting the SNN model with a good memory-accuracy trade-off at the end. Therefore, the memory saving offered by these state-of-the-art to meet the targeted accuracy is limited, thereby hindering processing SNNs on the resource-constrained systems (e.g., the IoT-Edge devices). Towards this, we propose Q-SpiNN, a novel quantization framework for memory-efficient SNNs. The key mechanisms of the Q-SpiNN are: (1) employing quantization for different SNN parameters based on their significance to the accuracy, (2) exploring different combinations of quantization schemes, precision levels, and rounding schemes to find efficient SNN model candidates, and (3) developing an algorithm that quantifies the benefit of the memory-accuracy trade-off obtained by the candidates, and selects the Pareto-optimal one. The experimental results show that, for the unsupervised network, the Q-SpiNN reduces the memory footprint by ca. 4x, while maintaining the accuracy within 1% from the baseline on the MNIST dataset. For the supervised network, the Q-SpiNN reduces the memory by ca. 2x, while keeping the accuracy within 2% from the baseline on the DVS-Gesture dataset.
We implement training of neural networks in secure multi-party computation (MPC) using quantization commonly used in said setting. We are the first to present an MNIST classifier purely trained in MPC that comes within 0.2 percent of the accuracy of the same convolutional neural network trained via plaintext computation. More concretely, we have trained a network with two convolutional and two dense layers to 99.2% accuracy in 3.5 hours (under one hour for 99% accuracy). We have also implemented AlexNet for CIFAR-10, which converges in a few hours. We develop novel protocols for exponentiation and inverse square root. Finally, we present experiments in a range of MPC security models for up to ten parties, both with honest and dishonest majority as well as semi-honest and malicious security.
Existing deep quantization methods provided an efficient solution for large-scale image retrieval. However, the significant intra-class variations like pose, illumination, and expressions in face images, still pose a challenge for face image retrieval. In light of this, face image retrieval requires sufficiently powerful learning metrics, which are absent in current deep quantization works. Moreover, to tackle the growing unseen identities in the query stage, face image retrieval drives more demands regarding model generalization and system scalability than general image retrieval tasks. This paper integrates product quantization with orthonormal constraints into an end-to-end deep learning framework to effectively retrieve face images. Specifically, a novel scheme that uses predefined orthonormal vectors as codewords is proposed to enhance the quantization informativeness and reduce codewords' redundancy. A tailored loss function maximizes discriminability among identities in each quantization subspace for both the quantized and original features. An entropy-based regularization term is imposed to reduce the quantization error. Experiments are conducted on four commonly-used face datasets under both seen and unseen identities retrieval settings. Our method outperforms all the compared deep hashing/quantization state-of-the-arts under both settings. Results validate the effectiveness of the proposed orthonormal codewords in improving models' standard retrieval performance and generalization ability. Combing with further experiments on two general image datasets, it demonstrates the broad superiority of our method for scalable image retrieval.
As edge devices become prevalent, deploying Deep Neural Networks (DNN) on edge devices has become a critical issue. However, DNN requires a high computational resource which is rarely available for edge devices. To handle this, we propose a novel model compression method for the devices with limited computational resources, called PQK consisting of pruning, quantization, and knowledge distillation (KD) processes. Unlike traditional pruning and KD, PQK makes use of unimportant weights pruned in the pruning process to make a teacher network for training a better student network without pre-training the teacher model. PQK has two phases. Phase 1 exploits iterative pruning and quantization-aware training to make a lightweight and power-efficient model. In phase 2, we make a teacher network by adding unimportant weights unused in phase 1 to a pruned network. By using this teacher network, we train the pruned network as a student network. In doing so, we do not need a pre-trained teacher network for the KD framework because the teacher and the student networks coexist within the same network. We apply our method to the recognition model and verify the effectiveness of PQK on keyword spotting (KWS) and image recognition.
Representing deep neural networks (DNNs) in low-precision is a promising approach to enable efficient acceleration and memory reduction. Previous methods that train DNNs in low-precision typically keep a copy of weights in high-precision during the weight updates. Directly training with low-precision weights leads to accuracy degradation due to complex interactions between the low-precision number systems and the learning algorithms. To address this issue, we develop a co-designed low-precision training framework, termed LNS-Madam, in which we jointly design a logarithmic number system (LNS) and a multiplicative weight update algorithm (Madam). We prove that LNS-Madam results in low quantization error during weight updates, leading to stable performance even if the precision is limited. We further propose a hardware design of LNS-Madam that resolves practical challenges in implementing an efficient datapath for LNS computations. Our implementation effectively reduces energy overhead incurred by LNS-to-integer conversion and partial sum accumulation. Experimental results show that LNS-Madam achieves comparable accuracy to full-precision counterparts with only 8 bits on popular computer vision and natural language tasks. Compared to FP32 and FP8, LNS-Madam reduces the energy consumption by over 90% and 55%, respectively.
Recently, transformer has achieved remarkable performance on a variety of computer vision applications. Compared with mainstream convolutional neural networks, vision transformers are often of sophisticated architectures for extracting powerful feature representations, which are more difficult to be developed on mobile devices. In this paper, we present an effective post-training quantization algorithm for reducing the memory storage and computational costs of vision transformers. Basically, the quantization task can be regarded as finding the optimal low-bit quantization intervals for weights and inputs, respectively. To preserve the functionality of the attention mechanism, we introduce a ranking loss into the conventional quantization objective that aims to keep the relative order of the self-attention results after quantization. Moreover, we thoroughly analyze the relationship between quantization loss of different layers and the feature diversity, and explore a mixed-precision quantization scheme by exploiting the nuclear norm of each attention map and output feature. The effectiveness of the proposed method is verified on several benchmark models and datasets, which outperforms the state-of-the-art post-training quantization algorithms. For instance, we can obtain an 81.29\% top-1 accuracy using DeiT-B model on ImageNet dataset with about 8-bit quantization.
Generally speaking, the main objective when training a neural speech synthesis system is to synthesize natural and expressive speech from the output layer of the neural network without much attention given to the hidden layers. However, by learning useful latent representation, the system can be used for many more practical scenarios. In this paper, we investigate the use of quantized vectors to model the latent linguistic embedding and compare it with the continuous counterpart. By enforcing different policies over the latent spaces in the training, we are able to obtain a latent linguistic embedding that takes on different properties while having a similar performance in terms of quality and speaker similarity. Our experiments show that the voice cloning system built with vector quantization has only a small degradation in terms of perceptive evaluations, but has a discrete latent space that is useful for reducing the representation bit-rate, which is desirable for data transferring, or limiting the information leaking, which is important for speaker anonymization and other tasks of that nature.