We introduce a quantization-aware training algorithm that guarantees avoiding numerical overflow when reducing the precision of accumulators during inference. We leverage weight normalization as a means of constraining parameters during training using accumulator bit width bounds that we derive. We evaluate our algorithm across multiple quantized models that we train for different tasks, showing that our approach can reduce the precision of accumulators while maintaining model accuracy with respect to a floating-point baseline. We then show that this reduction translates to increased design efficiency for custom FPGA-based accelerators. Finally, we show that our algorithm not only constrains weights to fit into an accumulator of user-defined bit width, but also increases the sparsity and compressibility of the resulting weights. Across all of our benchmark models trained with 8-bit weights and activations, we observe that constraining the hidden layers of quantized neural networks to fit into 16-bit accumulators yields an average 98.2% sparsity with an estimated compression rate of 46.5x all while maintaining 99.2% of the floating-point performance.
相關內容
We present a method for fast 3D reconstruction and real-time rendering of dynamic humans from monocular videos with accompanying parametric body fits. Our method can reconstruct a dynamic human in less than 3h using a single GPU, compared to recent state-of-the-art alternatives that take up to 72h. These speedups are obtained by using a lightweight deformation model solely based on linear blend skinning, and an efficient factorized volumetric representation for modeling the shape and color of the person in canonical pose. Moreover, we propose a novel local ray marching rendering which, by exploiting standard GPU hardware and without any baking or conversion of the radiance field, allows visualizing the neural human on a mobile VR device at 40 frames per second with minimal loss of visual quality. Our experimental evaluation shows superior or competitive results with state-of-the art methods while obtaining large training speedup, using a simple model, and achieving real-time rendering.
It is a widely observed phenomenon in nonparametric statistics that rate-optimal estimators balance bias and stochastic error. The recent work on overparametrization raises the question whether rate-optimal estimators exist that do not obey this trade-off. In this work we consider pointwise estimation in the Gaussian white noise model with $\beta$-H\"older smooth regression function f. It is shown that an estimator with worst-case bias $\lesssim n^{-\beta/(2\beta+1)}=: \psi_n$ must necessarily also have a worst-case mean absolute deviation that is lower bounded by $\gtrsim \psi_n.$ This proves that any estimator achieving the minimax optimal pointwise estimation rate $\psi_n$ must necessarily balance worst-case bias and worst-case mean absolute deviation. To derive the result, we establish an abstract inequality relating the change of expectation for two probability measures to the mean absolute deviation.
Optical Image Stabilization (OIS) system in mobile devices reduces image blurring by steering lens to compensate for hand jitters. However, OIS changes intrinsic camera parameters (i.e. $\mathrm{K}$ matrix) dynamically which hinders accurate camera pose estimation or 3D reconstruction. Here we propose a novel neural network-based approach that estimates $\mathrm{K}$ matrix in real-time so that pose estimation or scene reconstruction can be run at camera native resolution for the highest accuracy on mobile devices. Our network design takes gratified projection model discrepancy feature and 3D point positions as inputs and employs a Multi-Layer Perceptron (MLP) to approximate $f_{\mathrm{K}}$ manifold. We also design a unique training scheme for this network by introducing a Back propagated PnP (BPnP) layer so that reprojection error can be adopted as the loss function. The training process utilizes precise calibration patterns for capturing accurate $f_{\mathrm{K}}$ manifold but the trained network can be used anywhere. We name the proposed Dynamic Intrinsic Manifold Estimation network as DIME-Net and have it implemented and tested on three different mobile devices. In all cases, DIME-Net can reduce reprojection error by at least $64\%$ indicating that our design is successful.
Many empirical studies have demonstrated the performance benefits of conditional computation in neural networks, including reduced inference time and power consumption. We study the fundamental limits of neural conditional computation from the perspective of memorization capacity. For Rectified Linear Unit (ReLU) networks without conditional computation, it is known that memorizing a collection of $n$ input-output relationships can be accomplished via a neural network with $O(\sqrt{n})$ neurons. Calculating the output of this neural network can be accomplished using $O(\sqrt{n})$ elementary arithmetic operations of additions, multiplications and comparisons for each input. Using a conditional ReLU network, we show that the same task can be accomplished using only $O(\log n)$ operations per input. This represents an almost exponential improvement as compared to networks without conditional computation. We also show that the $\Theta(\log n)$ rate is the best possible. Our achievability result utilizes a general methodology to synthesize a conditional network out of an unconditional network in a computationally-efficient manner, bridging the gap between unconditional and conditional architectures.
Social ambiance describes the context in which social interactions happen, and can be measured using speech audio by counting the number of concurrent speakers. This measurement has enabled various mental health tracking and human-centric IoT applications. While on-device Socal Ambiance Measure (SAM) is highly desirable to ensure user privacy and thus facilitate wide adoption of the aforementioned applications, the required computational complexity of state-of-the-art deep neural networks (DNNs) powered SAM solutions stands at odds with the often constrained resources on mobile devices. Furthermore, only limited labeled data is available or practical when it comes to SAM under clinical settings due to various privacy constraints and the required human effort, further challenging the achievable accuracy of on-device SAM solutions. To this end, we propose a dedicated neural architecture search framework for Energy-efficient and Real-time SAM (ERSAM). Specifically, our ERSAM framework can automatically search for DNNs that push forward the achievable accuracy vs. hardware efficiency frontier of mobile SAM solutions. For example, ERSAM-delivered DNNs only consume 40 mW x 12 h energy and 0.05 seconds processing latency for a 5 seconds audio segment on a Pixel 3 phone, while only achieving an error rate of 14.3% on a social ambiance dataset generated by LibriSpeech. We can expect that our ERSAM framework can pave the way for ubiquitous on-device SAM solutions which are in growing demand.
Deep neural networks provide state-of-the-art accuracy for vision tasks but they require significant resources for training. Thus, they are trained on cloud servers far from the edge devices that acquire the data. This issue increases communication cost, runtime and privacy concerns. In this study, a novel hierarchical training method for deep neural networks is proposed that uses early exits in a divided architecture between edge and cloud workers to reduce the communication cost, training runtime and privacy concerns. The method proposes a brand-new use case for early exits to separate the backward pass of neural networks between the edge and the cloud during the training phase. We address the issues of most available methods that due to the sequential nature of the training phase, cannot train the levels of hierarchy simultaneously or they do it with the cost of compromising privacy. In contrast, our method can use both edge and cloud workers simultaneously, does not share the raw input data with the cloud and does not require communication during the backward pass. Several simulations and on-device experiments for different neural network architectures demonstrate the effectiveness of this method. It is shown that the proposed method reduces the training runtime by 29% and 61% in CIFAR-10 classification experiment for VGG-16 and ResNet-18 when the communication with the cloud is done at a low bit rate channel. This gain in the runtime is achieved whilst the accuracy drop is negligible. This method is advantageous for online learning of high-accuracy deep neural networks on low-resource devices such as mobile phones or robots as a part of an edge-cloud system, making them more flexible in facing new tasks and classes of data.
Graph Neural Networks (GNNs) with attention have been successfully applied for learning visual feature matching. However, current methods learn with complete graphs, resulting in a quadratic complexity in the number of features. Motivated by a prior observation that self- and cross- attention matrices converge to a sparse representation, we propose ClusterGNN, an attentional GNN architecture which operates on clusters for learning the feature matching task. Using a progressive clustering module we adaptively divide keypoints into different subgraphs to reduce redundant connectivity, and employ a coarse-to-fine paradigm for mitigating miss-classification within images. Our approach yields a 59.7% reduction in runtime and 58.4% reduction in memory consumption for dense detection, compared to current state-of-the-art GNN-based matching, while achieving a competitive performance on various computer vision tasks.
High-performance deep neural network (DNN)-based systems are in high demand in edge environments. Due to its high computational complexity, it is challenging to deploy DNNs on edge devices with strict limitations on computational resources. In this paper, we derive a compact while highly-accurate DNN model, termed dsODENet, by combining recently-proposed parameter reduction techniques: Neural ODE (Ordinary Differential Equation) and DSC (Depthwise Separable Convolution). Neural ODE exploits a similarity between ResNet and ODE, and shares most of weight parameters among multiple layers, which greatly reduces the memory consumption. We apply dsODENet to a domain adaptation as a practical use case with image classification datasets. We also propose a resource-efficient FPGA-based design for dsODENet, where all the parameters and feature maps except for pre- and post-processing layers can be mapped onto on-chip memories. It is implemented on Xilinx ZCU104 board and evaluated in terms of domain adaptation accuracy, inference speed, FPGA resource utilization, and speedup rate compared to a software counterpart. The results demonstrate that dsODENet achieves comparable or slightly better domain adaptation accuracy compared to our baseline Neural ODE implementation, while the total parameter size without pre- and post-processing layers is reduced by 54.2% to 79.8%. Our FPGA implementation accelerates the inference speed by 23.8 times.
Sampling methods (e.g., node-wise, layer-wise, or subgraph) has become an indispensable strategy to speed up training large-scale Graph Neural Networks (GNNs). However, existing sampling methods are mostly based on the graph structural information and ignore the dynamicity of optimization, which leads to high variance in estimating the stochastic gradients. The high variance issue can be very pronounced in extremely large graphs, where it results in slow convergence and poor generalization. In this paper, we theoretically analyze the variance of sampling methods and show that, due to the composite structure of empirical risk, the variance of any sampling method can be decomposed into \textit{embedding approximation variance} in the forward stage and \textit{stochastic gradient variance} in the backward stage that necessities mitigating both types of variance to obtain faster convergence rate. We propose a decoupled variance reduction strategy that employs (approximate) gradient information to adaptively sample nodes with minimal variance, and explicitly reduces the variance introduced by embedding approximation. We show theoretically and empirically that the proposed method, even with smaller mini-batch sizes, enjoys a faster convergence rate and entails a better generalization compared to the existing methods.
Deep convolutional neural networks (CNNs) have recently achieved great success in many visual recognition tasks. However, existing deep neural network models are computationally expensive and memory intensive, hindering their deployment in devices with low memory resources or in applications with strict latency requirements. Therefore, a natural thought is to perform model compression and acceleration in deep networks without significantly decreasing the model performance. During the past few years, tremendous progress has been made in this area. In this paper, we survey the recent advanced techniques for compacting and accelerating CNNs model developed. These techniques are roughly categorized into four schemes: parameter pruning and sharing, low-rank factorization, transferred/compact convolutional filters, and knowledge distillation. Methods of parameter pruning and sharing will be described at the beginning, after that the other techniques will be introduced. For each scheme, we provide insightful analysis regarding the performance, related applications, advantages, and drawbacks etc. Then we will go through a few very recent additional successful methods, for example, dynamic capacity networks and stochastic depths networks. After that, we survey the evaluation matrix, the main datasets used for evaluating the model performance and recent benchmarking efforts. Finally, we conclude this paper, discuss remaining challenges and possible directions on this topic.
小貼士
登錄享
注冊地址： 北京市海淀區羊坊店路18號2幢3層301-191