Recent studies have shown that neural network (NN) based image classifiers are highly vulnerable to adversarial examples, which poses a threat to security-sensitive image recognition task. Prior work has shown that JPEG compression can combat the drop in classification accuracy on adversarial examples to some extent. But, as the compression ratio increases, traditional JPEG compression is insufficient to defend those attacks but can cause an abrupt accuracy decline to the benign images. In this paper, with the aim of fully filtering the adversarial perturbations, we firstly make modifications to traditional JPEG compression algorithm which becomes more favorable for NN. Specifically, based on an analysis of the frequency coefficient, we design a NN-favored quantization table for compression. Considering compression as a data augmentation strategy, we then combine our model-agnostic preprocess with noisy training. We fine-tune the pre-trained model by training with images encoded at different compression levels, thus generating multiple classifiers. Finally, since lower (higher) compression ratio can remove both perturbations and original features slightly (aggressively), we use these trained multiple models for model ensemble. The majority vote of the ensemble of models is adopted as final predictions. Experiments results show our method can improve defense efficiency while maintaining original accuracy.
To achieve higher accuracy in machine learning tasks, very deep convolutional neural networks (CNNs) are designed recently. However, the large memory access of deep CNNs will lead to high power consumption. A variety of hardware-friendly compression methods have been proposed to reduce the data transfer bandwidth by exploiting the sparsity of feature maps. Most of them focus on designing a specialized encoding format to increase the compression ratio. Differently, we observe and exploit the sparsity distinction between activations in earlier and later layers to improve the compression ratio. We propose a novel hardware-friendly transform-based method named 1D-Discrete Cosine Transform on Channel dimension with Masks (DCT-CM), which intelligently combines DCT, masks, and a coding format to compress activations. The proposed algorithm achieves an average compression ratio of 2.9x (53% higher than the state-of-the-art transform-based feature map compression works) during inference on ResNet-50 with an 8-bit quantization scheme.
Pre-trained language models like Ernie or Bert are currently used in many applications. These models come with a set of pre-trained weights typically obtained in unsupervised/self-supervised modality on a huge amount of data. After that, they are fine-tuned on a specific task. Applications then use these models for inference, and often some additional constraints apply, like low power-budget or low latency between input and output. The main avenue to meet these additional requirements for the inference settings, is to use low precision computation (e.g. INT8 rather than FP32), but this comes with a cost of deteriorating the functional performance (e.g. accuracy) of the model. Some approaches have been developed to tackle the problem and go beyond the limitations of the PTO (Post-Training Quantization), more specifically the QAT (Quantization Aware Training, see [4]) is a procedure that interferes with the training process in order to make it affected (or simply disturbed) by the quantization phase during the training itself. Besides QAT, recently Intel-Habana Labs have proposed an additional and more direct way to make the training results more robust to subsequent quantization which uses a regularizer, therefore changing the loss function that drives the training procedure. But their proposal does not work out-of-the-box for pre-trained models like Ernie, for example. In this short paper we show why this is not happening (for the Ernie case) and we propose a very basic way to deal with it, sharing as well some initial results (increase in final INT8 accuracy) that might be of interest to practitioners willing to use Ernie in their applications, in low precision regime.
Over the years, accelerating neural networks with quantization has been widely studied. Unfortunately, prior efforts with diverse precisions (e.g., 1-bit weights and 2-bit activations) are usually restricted by limited precision support on GPUs (e.g., int1 and int4). To break such restrictions, we introduce the first Arbitrary Precision Neural Network framework (APNN-TC) to fully exploit quantization benefits on Ampere GPU Tensor Cores. Specifically, APNN-TC first incorporates a novel emulation algorithm to support arbitrary short bit-width computation with int1 compute primitives and XOR/AND Boolean operations. Second, APNN-TC integrates arbitrary precision layer designs to efficiently map our emulation algorithm to Tensor Cores with novel batching strategies and specialized memory organization. Third, APNN-TC embodies a novel arbitrary precision NN design to minimize memory access across layers and further improve performance. Extensive evaluations show that APNN-TC can achieve significant speedup over CUTLASS kernels and various NN models, such as ResNet and VGG.
Over-the-air computation (AirComp) has recently been recognized as a promising scheme for a fusion center to achieve fast distributed data aggregation in wireless networks via exploiting the superposition property of multiple-access channels. Since it is challenging to provide reliable data aggregation for a large number of devices using AirComp, in this paper, we propose to enable AirComp via the cloud radio access network (Cloud-RAN) architecture, where a large number of antennas are deployed at separate sites called remote radio heads (RRHs). However, the potential densification gain provided by Cloud-RAN is generally bottlenecked by the limited capacity of the fronthaul links connecting the RRHs and the fusion center. To this end, we formulate a joint design problem for AirComp transceivers and quantization bits allocation and propose an efficient algorithm to tackle this problem. Our numerical results shows the advantages of the proposed architecture compared with the state-of-the-art solutions.
We analyze the symbol error probability (SEP) of $M$-ary pulse amplitude modulation ($M$-PAM) receivers equipped with optimal low-resolution quantizers. We first show that the optimum detector can be reduced to a simple decision rule. Using this simplification, an exact SEP expression for quantized $M$-PAM receivers is obtained when Nakagami-$m$ fading channel is considered. The derived expression enables the optimization of the quantizer and/or constellation under the minimum SEP criterion. Our analysis of optimal quantization for equidistant $M$-PAM receiver reveals the existence of error floor which decays at a double exponential rate with increasing quantization bits, $b$. Moreover, by also allowing the transmitter to optimize the constellation based on the statistics of the fading channel, we prove that the error floor can be eliminated but at a lower decay exponent than the unquantized case. Characterization of this decay exponent is provided in this paper. We also expose the outage performance limitations of SEP-optimal uniform quantizers. To be more precise, its decay exponent does not improve with $b$. Lastly, we demonstrate that the decay exponent of a quantized receiver can be complemented by receive antenna diversity techniques.
This paper develops an efficient procedure for designing low-complexity codebooks for precoding in a full-dimension (FD) multiple-input multiple-output (MIMO) system with a uniform planar array (UPA) antenna at the transmitter (Tx) using tensor learning. In particular, instead of using statistical channel models, we utilize a model-free data-driven approach with foundations in machine learning to generate codebooks that adapt to the surrounding propagation conditions. We use a tensor representation of the FD-MIMO channel and exploit its properties to design quantized version of the channel precoders. We find the best representation of the optimal precoder as a function of Kronecker Product (KP) of two low-dimensional precoders, respectively corresponding to the horizontal and vertical dimensions of the UPA, obtained from the tensor decomposition of the channel. We then quantize this precoder to design product codebooks such that an average loss in mutual information due to quantization of channel state information (CSI) is minimized. The key technical contribution lies in exploiting the constraints on the precoders to reduce the product codebook design problem to an unsupervised clustering problem on a Cartesian Product Grassmann manifold (CPM), where the cluster centroids form a finite-sized precoder codebook. This codebook can be found efficiently by running a $K$-means clustering on the CPM. With a suitable induced distance metric on the CPM, we show that the construction of product codebooks is equivalent to finding the optimal set of centroids on the factor manifolds corresponding to the horizontal and vertical dimensions. Simulation results are presented to demonstrate the capability of the proposed design criterion in learning the codebooks and the attractive performance of the designed codebooks.
Large machine learning models achieve unprecedented performance on various tasks and have evolved as the go-to technique. However, deploying these compute and memory hungry models on resource constraint environments poses new challenges. In this work, we propose mathematically provable Representer Sketch, a concise set of count arrays that can approximate the inference procedure with simple hashing computations and aggregations. Representer Sketch builds upon the popular Representer Theorem from kernel literature, hence the name, providing a generic fundamental alternative to the problem of efficient inference that goes beyond the popular approach such as quantization, iterative pruning and knowledge distillation. A neural network function is transformed to its weighted kernel density representation, which can be very efficiently estimated with our sketching algorithm. Empirically, we show that Representer Sketch achieves up to 114x reduction in storage requirement and 59x reduction in computation complexity without any drop in accuracy.
We study the Witten-type topological field theory(W-TFT) of self-organized criticality(SOC) for stochastic neural networks. The Parisi-Sourlas-Wu quantization of general stochastic differential equations (SDEs) for neural networks, the Becchi-Rouet-Stora-Tyutin(BRST)-symmetry of the diffusion system and the relation between spontaneous breaking and instantons connecting steady states of the SDEs, as well as the sufficient and necessary condition on pseudo-supersymmetric stochastic neural networks are obtained. Suppose neuronal avalanche is a mechanism of cortical information processing and storage \cite{Beggs}\cite{Plenz1}\cite{Plenz2} and the model of stochastic neural networks\cite{Dayan} is correct, as well as the SOC system can be looked upon as a W-TFT with spontaneously broken BRST symmetry. Then we should recover the neuronal avalanches and spontaneously broken BRST symmetry from the model of stochastic neural networks. We find that, provided the divergence of drift coefficients is small and non-constant, the model of stochastic neural networks is BRST symmetric. That is, if the SOC of brain neural networks system can be looked upon as a W-TFT with spontaneously broken BRST symmetry, then the general model of stochastic neural networks which be extensively used in neuroscience \cite{Dayan} is not enough to describe the SOC. On the other hand, using the Fokker-Planck equation, we show the sufficient condition on diffusion so that there exists a steady state probability distribution for the stochastic neural networks. Rhythms of the firing rates of the neuronal networks arise from the process, meanwhile some biological laws are conserved.
The spiking neural network (SNN) has been attracting considerable attention not only as a mathematical model for the brain, but also as an energy-efficient information processing model for real-world applications. In particular, SNNs based on temporal coding are expected to be much more efficient than those based on rate coding, because the former requires substantially fewer spikes to carry out tasks. As SNNs are continuous-state and continuous-time models, it is favorable to implement them with analog VLSI circuits. However, the construction of the entire system with continuous-time analog circuits would be infeasible when the system size is very large. Therefore, mixed-signal circuits must be employed, and the time discretization and quantization of the synaptic weights are necessary. Moreover, the analog VLSI implementation of SNNs exhibits non-idealities, such as the effects of noise and device mismatches, as well as other constraints arising from the analog circuit operation. In this study, we investigated the effects of the time discretization and/or weight quantization on the performance of SNNs. Furthermore, we elucidated the effects the lower bound of the membrane potentials and the temporal fluctuation of the firing threshold. Finally, we propose an optimal approach for the mapping of mathematical SNN models to analog circuits with discretized time.
One-shot voice conversion (VC), which performs conversion across arbitrary speakers with only a single target-speaker utterance for reference, can be effectively achieved by speech representation disentanglement. Existing work generally ignores the correlation between different speech representations during training, which causes leakage of content information into the speaker representation and thus degrades VC performance. To alleviate this issue, we employ vector quantization (VQ) for content encoding and introduce mutual information (MI) as the correlation metric during training, to achieve proper disentanglement of content, speaker and pitch representations, by reducing their inter-dependencies in an unsupervised manner. Experimental results reflect the superiority of the proposed method in learning effective disentangled speech representations for retaining source linguistic content and intonation variations, while capturing target speaker characteristics. In doing so, the proposed approach achieves higher speech naturalness and speaker similarity than current state-of-the-art one-shot VC systems. Our code, pre-trained models and demo are available at https://github.com/Wendison/VQMIVC.
A significant amount of work has been done on adversarial attacks that inject imperceptible noise to images to deteriorate the image classification performance of deep models. However, most of the existing studies consider attacks in the digital (pixel) domain where an image acquired by an image sensor with sampling and quantization has been recorded. This paper, for the first time, introduces an optical adversarial attack, which physically alters the light field information arriving at the image sensor so that the classification model yields misclassification. More specifically, we modulate the phase of the light in the Fourier domain using a spatial light modulator placed in the photographic system. The operative parameters of the modulator are obtained by gradient-based optimization to maximize cross-entropy and minimize distortions. We present experiments based on both simulation and a real hardware optical system, from which the feasibility of the proposed optical attack is demonstrated. It is also verified that the proposed attack is completely different from common optical-domain distortions such as spherical aberration, defocus, and astigmatism in terms of both perturbation patterns and classification results.
A change of the prevalent supervised learning techniques is foreseeable in the near future: from the complex, computational expensive algorithms to more flexible and elementary training ones. The strong revitalization of randomized algorithms can be framed in this prospect steering. We recently proposed a model for distributed classification based on randomized neural networks and hyperdimensional computing, which takes into account cost of information exchange between agents using compression. The use of compression is important as it addresses the issues related to the communication bottleneck, however, the original approach is rigid in the way the compression is used. Therefore, in this work, we propose a more flexible approach to compression and compare it to conventional compression algorithms, dimensionality reduction, and quantization techniques.
Machine learning algorithms deployed on edge devices must meet certain resource constraints and efficiency requirements. Random Vector Functional Link (RVFL) networks are favored for such applications due to their simple design and training efficiency. We propose a modified RVFL network that avoids computationally expensive matrix operations during training, thus expanding the network's range of potential applications. Our modification replaces the least-squares classifier with the Generalized Learning Vector Quantization (GLVQ) classifier, which only employs simple vector and distance calculations. The GLVQ classifier can also be considered an improvement upon certain classification algorithms popularly used in the area of Hyperdimensional Computing. The proposed approach achieved state-of-the-art accuracy on a collection of datasets from the UCI Machine Learning Repository - higher than previously proposed RVFL networks. We further demonstrate that our approach still achieves high accuracy while severely limited in training iterations (using on average only 21% of the least-squares classifier computational costs).
This study exploits previously demonstrated properties such as sensitivity to the spatial extent and the intensity of local image contrast of the quantization error in the output of a Self Organizing Map (SOM QE). Here, the SOM QE is applied to double color staining based cell viability data in 96 image simulations. The results show that the SOM QE consistently and in only a few seconds detects fine regular spatial increases in relative amounts of RED or GREEN pixel staining across the test images, reflecting small, systematic increases or decreases in the percentage of theoretical cell viability below the critical threshold. Such small changes may carry clinical significance, but are almost impossible to detect by human vision. Moreover, we demonstrate a clear sensitivity of the SOM QE to differences in the relative physical luminance (Y) of the colors, which here translates into a RED GREEN color selectivity. Across differences in relative luminance, the SOM QE exhibits consistently greater sensitivity to the smallest spatial increases in RED image pixels compared with smallest increases of identical spatial extents in GREEN image pixels. Further selective color contrast studies on simulations of biological imaging data will allow generating increasingly larger benchmark datasets and, ultimately, unravel the full potential of fast, economic, and unprecedentedly precise biological data analysis using the SOM QE.
Federated learning (FL) has been recognized as a viable distributed learning paradigm which trains a machine learning model collaboratively with massive mobile devices in the wireless edge while protecting user privacy. Although various communication schemes have been proposed to expedite the FL process, most of them have assumed ideal wireless channels which provide reliable and lossless communication links between the server and mobile clients. Unfortunately, in practical systems with limited radio resources such as constraint on the training latency and constraints on the transmission power and bandwidth, transmission of a large number of model parameters inevitably suffers from quantization errors (QE) and transmission outage (TO). In this paper, we consider such non-ideal wireless channels, and carry out the first analysis showing that the FL convergence can be severely jeopardized by TO and QE, but intriguingly can be alleviated if the clients have uniform outage probabilities. These insightful results motivate us to propose a robust FL scheme, named FedTOE, which performs joint allocation of wireless resources and quantization bits across the clients to minimize the QE while making the clients have the same TO probability. Extensive experimental results are presented to show the superior performance of FedTOE for deep learning-based classification tasks with transmission latency constraints.
Quantization is used to speed up execution time and save power when runnning Deep neural networks (DNNs) on edge devices like AI chips. To investigate the effect of quantization, we need performing inference after quantizing the weights of DNN with 32-bit floating-point precision by a some bit width, and then quantizing them back to 32-bit floating-point precision. This is because the DNN library can only handle floating-point numbers. However, the accuracy of the emulation does not provide accurate precision. We need accurate precision to detect overflow in MAC operations or to verify the operation on edge de vices. We have developed PyParch, a DNN library that executes quantized DNNs (QNNs) with exactly the same be havior as hardware. In this paper, we describe a new proposal and implementation of PyParch. As a result of the evaluation, the accuracy of QNNs with arbitrary bit widths can be estimated for la rge and complex DNNs such as YOLOv5, and the overflow can be detected. We evaluated the overhead of the emulation time and found that it was 5.6 times slower for QNN and 42 times slower for QNN with overflow detection compared to the normal DNN execution time.
The spectacular success of deep generative models calls for quantitative tools to measure their statistical performance. Divergence frontiers have recently been proposed as an evaluation framework for generative models, due to their ability to measure the quality-diversity trade-off inherent to deep generative modeling. We establish non-asymptotic bounds on the sample complexity of divergence frontiers. We also introduce frontier integrals which provide summary statistics of divergence frontiers. We show how smoothed estimators such as Good-Turing or Krichevsky-Trofimov can overcome the missing mass problem and lead to faster rates of convergence. We illustrate the theoretical results with numerical examples from natural language processing and computer vision.
While neural networks have advanced the frontiers in many applications, they often come at a high computational cost. Reducing the power and latency of neural network inference is key if we want to integrate modern networks into edge devices with strict power and compute requirements. Neural network quantization is one of the most effective ways of achieving these savings but the additional noise it induces can lead to accuracy degradation. In this white paper, we introduce state-of-the-art algorithms for mitigating the impact of quantization noise on the network's performance while maintaining low-bit weights and activations. We start with a hardware motivated introduction to quantization and then consider two main classes of algorithms: Post-Training Quantization (PTQ) and Quantization-Aware-Training (QAT). PTQ requires no re-training or labelled data and is thus a lightweight push-button approach to quantization. In most cases, PTQ is sufficient for achieving 8-bit quantization with close to floating-point accuracy. QAT requires fine-tuning and access to labeled training data but enables lower bit quantization with competitive results. For both solutions, we provide tested pipelines based on existing literature and extensive experimentation that lead to state-of-the-art performance for common deep learning models and tasks.
We present an algorithm for searching image collections using free-hand sketches that describe the appearance and relative positions of multiple objects. Sketch based image retrieval (SBIR) methods predominantly match queries containing a single, dominant object invariant to its position within an image. Our work exploits drawings as a concise and intuitive representation for specifying entire scene compositions. We train a convolutional neural network (CNN) to encode masked visual features from sketched objects, pooling these into a spatial descriptor encoding the spatial relationships and appearances of objects in the composition. Training the CNN backbone as a Siamese network under triplet loss yields a metric search embedding for measuring compositional similarity which may be efficiently leveraged for visual search by applying product quantization.