Batch Normalization and Dropout Regularization in Training Deep Neural Networks with Label Noise

Batch Normalization (BN) techniques have been proposed to reduce the so-called Internal Covariate Shift (ICS) by attempting to keep the distributions of layer outputs unchanged. Experiments have shown their effectiveness on training deep neural networks. However, since only the first two moments are controlled in these BN techniques, it seems that a weak constraint is imposed on layer distributions and furthermore whether such constraint can reduce ICS is unknown. Thus this paper proposes a measure for ICS by using the Earth Mover (EM) distance and then derives the upper and lower bounds for the measure to provide a theoretical analysis of BN. The upper bound has shown that BN techniques can control ICS only for the outputs with low dimensions and small noise whereas their control is not effective in other cases. This paper also proves that such control is just a bounding of ICS rather than a reduction of ICS. Meanwhile, the analysis shows that the high-order moments and noise, which BN cannot control, have great impact on the lower bound. Based on such analysis, this paper furthermore proposes an algorithm that unitizes the outputs with an adjustable parameter to further bound ICS in order to cope with the problems of BN. The upper bound for the proposed unitization is noise-free and only dominated by the parameter. Thus, the parameter can be trained to tune the bound and further to control ICS. Besides, the unitization is embedded into the framework of BN to reduce the information loss. The experiments show that this proposed algorithm outperforms existing BN techniques on CIFAR-10, CIFAR-100 and ImageNet datasets.

Batch Normalization (BN) is a well-known technique used in training deep neural networks. The main idea behind batch normalization is to normalize the features of the layers (i.e., transforming them to have a mean equal to zero and a variance equal to one). Such a procedure encourages the optimization landscape of the loss function to be smoother, and improves the learning of the networks for both speed and performance. In this paper, we demonstrate that the performance of the network can be improved, if the distributions of the features of the output in the same layer are similar. As normalizing based on mean and variance does not necessarily make the features to have the same distribution, we propose a new normalization scheme: Batch Normalization with Skewness Reduction (BNSR). Comparing with other normalization approaches, BNSR transforms not just only the mean and variance, but also the skewness of the data. By tackling this property of a distribution, we are able to make the output distributions of the layers to be further similar. The nonlinearity of BNSR may further improve the expressiveness of the underlying network. Comparisons with other normalization schemes are tested on the CIFAR-100 and ImageNet datasets. Experimental results show that the proposed approach can outperform other state-of-the-arts that are not equipped with BNSR.

Recently deep-learning (DL) techniques have been widely adopted in side-channel power analysis. A DL-assisted SCA generally consists of two phases: a deep neural network (DNN) training phase and a follow-on attack phase using the trained DNN. However, currently the two phases are not well aligned, as there is no conclusion on what metric used in the training can result in the most effective attack in the second phase. When traditional loss functions such as negative log-likelihood (NLL) are used in training a DNN, the trained model does not yield optimal follow-on attack. Recently some information theoretical SCA leakage metrics are proposed, either as the validation metric to stop the DNN training with traditional loss functions, or as both the validation metric and the training loss function. None of those proposed metrics, however, directly measures the SCA effectiveness. We propose to conduct DNN training directly with a common SCA effectiveness metric, Guessing Entropy (GE). We overcome the prior practical difficulty of using GE in DNN training by utilizing the GEEA estimation algorithm introduced in CHES 2020. We show that using GEEA as either the validation metric or the loss function produces DNN models that lead to much more effective follow-on attacks. Our work consolidates the DL-assisted SCA framework with a consistent metric, which shows great potential to be adopted as the universal SCA-oriented DNN training framework.

An Entropic Optimal Transport loss for learning deep neural networks under label noise in remote sensing images

Deep Learning (DL) has become a prominent machine learning technique due to the availability of efficient computational resources in the form of Graphics Processing Units (GPUs), large-scale datasets and a variety of models. The newer generation of GPUs are being designed with special emphasis on optimizing performance for DL applications. Also, the availability of easy-to-use DL frameworks—like PyTorch and TensorFlow— has enhanced productivity of domain experts to work on their custom DL applications from diverse domains. However, existing Deep Neural Network (DNN) training approaches may not fully utilize the newly emerging powerful GPUs like the NVIDIA A100—this is the primary issue that we address in this paper. Our motivating analyses show that the GPU utilization on NVIDIA A100 can be as low as 43% using traditional DNN training approaches for small-to-medium DL models and input data size. This paper proposes AccDP—a data-parallel distributed DNN training approach—to accelerate GPU-based DL applications. AccDP exploits the Message Passing Interface (MPI) communication library coupled with the NVIDIA’s Multi-Process Service (MPS) to increase the amount of work assigned to parallel GPUs resulting in higher utilization of compute resources. We evaluate our proposed design on different small-to-medium DL models and input sizes on the state-of-the-art HPC clusters. By injecting more parallelism into DNN training using our approach, the evaluation shows up to 58% improvement in training performance on a single GPU and up to 62% on 16 GPUs compared to regular DNN training. Furthermore, we conduct an in-depth characterization to determine the impact of several DNN training factors and best practices—including the batch size and the number of data loading workers— to optimally utilize GPU devices. To the best of our knowledge, this is the first work that explores the use of MPS and MPI to maximize the utilization of GPUs in distributed DNN training.

Incorrectly labeled examples, or label noise, is common in real-world computer vision datasets. While the impact of label noise on learning in deep neural networks has been studied in prior work, these studies have exclusively focused on homogeneous label noise, i.e., the degree of label noise is the same across all categories. However, in the real-world, label noise is often heterogeneous, with some categories being affected to a greater extent than others. Here, we address this gap in the literature. We hypothesized that heterogeneous label noise would only affect the classes that had label noise unless there was transfer from those classes to the classes without label noise. To test this hypothesis, we designed a series of computer vision studies using MNIST, CIFAR-10, CIFAR-100, and MS-COCO where we imposed heterogeneous label noise during the training of multi-class, multi-task, and multi-label systems. Our results provide evidence in support of our hypothesis: label noise only affects the class affected by it unless there is transfer.

Batch normalization (BN) has been established as a very effective component in deep learning, largely helping accelerate the convergence of deep neural network (DNN) training. Nevertheless, its hardware architecture has not received much attention in the field of DNN on-device training processors. Several previous designs incur either high off-chip memory traffic or high circuit complexity, and hence have deficiencies in terms of hardware efficiency and performance. This article proposes approximately calculated BN (ACBN) to achieve a much better tradeoff between hardware efficiency and performance for DNN on-device training processors. The accuracy and convergence rate of the proposed ACBN have been extensively evaluated using four typical DNN models. Compared with the state-of-the-art reference design, the hardware simulation results show the proposed ACBN can at least reduce floating point operations by 22.2% and save external memory access by 33.3% on average. Moreover, the proposed ACBN introduces 63.6% data sparsity for the backward propagation of BN layers of VGG16 on average. To the best of our knowledge, we are the first to introduce data sparsity for the backward propagation of BN layers. The ACBN module is implemented on Zynq UltraScale <inline-formula xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink"> <tex-math notation="LaTeX">$+$</tex-math> </inline-formula> ZCU102 system-on-chip (SoC) field-programmable gate array (FPGA), and the results show that the implementation of ACBN hardware module saves 33.9% look-up table (LUT), 49.4% flip-flop (FF), 75% digital signal processor (DSP), and reduces the power by 12.4% compared with the reference design while achieving better performance.

Peer-to-peer deep learning with non-IID data

It has been shown that the learning rate is one of the most critical hyper-parameters for the overall performance of deep neural networks. In this paper, we propose a new method for setting the global learning rate, named random amplify learning rates (RALR), to improve the performance of any optimizer in training deep neural networks. Instead of monotonically decreasing the learning rate, we expect to escape saddle points or local minima by amplifying the learning rate between reasonable boundary values based on a given probability. Training with RALR rather than conventionally decreasing the learning rate achieves further improvement on networks’ performance without extra consumption. Remarkably, the RALR is complementary with state-of-the-art data augmentation and regularization methods. Besides, we empirically study its performance on image classification tasks, fine-grained classification tasks, object detection tasks, and machine translation tasks. Experiments demonstrate that RALR can bring a notable improvement while preventing overfitting when training deep neural networks. For example, the classification accuracy of ResNet-110 trained on the CIFAR-100 dataset using RALR achieves a 1.34% gain compared with ResNet-110 trained traditionally.

Deep neural networks (DNNs) have demonstrated effectiveness in many domains including object recognition, speech recognition, natural language processing, and health care. Typically, the computations involved in DNN training and inferencing are time consuming and require efficient implementations. Existing frameworks such as TensorFlow, Theano, Torch, Cognitive Tool Kit (CNTK), and Caffe enable Graphics Processing Unit (GPUs) as the status quo devices for DNN execution, leaving Central Processing Unit (CPUs) behind. Moreover, existing frameworks forgo or limit cross layer optimization opportunities that have the potential to improve performance by significantly reducing data movement through the memory hierarchy. In this article, we describe an alternative approach called SWIRL, a compiler that provides high-performance CPU implementations for DNNs. SWIRL is built on top of the existing domain-specific language (DSL) for DNNs called LATTE. SWIRL separates DNN specification and its schedule using predefined transformation recipes for tensors and layers commonly found in DNN layers. These recipes synergize with DSL constructs to generate high-quality fused, vectorized, and parallelized code for CPUs. On an Intel Xeon Platinum 8180M CPU, SWIRL achieves performance comparable with Tensorflow integrated with MKL-DNN; on average 1.00× of Tensorflow inference and 0.99× of Tensorflow training. It also outperforms the original LATTE compiler on average by 1.22× and 1.30× on inference and training, respectively.

Automatic modulation classification (AMC) has a wide range of applications in both civilian and military fields, such as industrial Internet of Things (IIoT) security, communication spectrum management, and military electronic countermeasures. However, label mislabeling often occurs in practical scenarios, significantly impacting the performance and robustness of deep neural networks (DNNs). In this article, we propose a meta-learning guided label noise distillation method to enhance the robustness of AMC models against label noise or errors. Specifically, we propose a teacher-student heterogeneous network (TSHN) to discriminate and distill label noise. Following the notion that labels represent information, a teacher network, utilizing trusted few-shot labeled samples, reevaluates and corrects labels for a considerable number of untrusted labeled samples through meta-learning. By dividing and conquering untrusted labeled samples according to their confidence levels, the student network learns more effectively. Additionally, we propose a multiview signal (MVS) method to further enhance the performance of hard-to-classify categories with few-shot trusted labeled samples. Extensive experiments on the RadioML2016 and HisarMod2019.1 data sets demonstrate that our methods significantly improve accuracy and robustness in signal AMC across diverse label noise scenarios, including symmetric, asymmetric, and mixed label noise. For example, compared to the baseline convolutional neural network with the cross-entropy loss, our proposed TSHN achieves a remarkable 1.26% to 36.84% accuracy improvement under symmetric label noise and 0.12% to 38.59% accuracy improvement under mixed label noise. Moreover, TSHN exhibits greater robustness to varying label noise rates compared to existing methods.

Survey on Deep Neural Networks in Speech and Vision Systems

Automated image classification is an essential task of the computer vision field. The tagging of images into a set of predefined groups is referred to as image classification. The implementation of computer vision to automate image classification would be beneficial because manual image evaluation and identification can be time-consuming, particularly when there are many images of different classes. Deep learning approaches are proven to overperform existing machine learning techniques in many fields in recent years, and computer vision is one of the most notable examples. The very deep neural network (VDCNN) is a powerful deep learning model for image classification, and this paper examines it briefly using MNIST handwritten digit dataset. This dataset is used to prove the efficacy of very deep neural networks over other deep learning models. The proposed study aims to comprehend the very deep neural network architecture used to accomplish a handwritten digit recognition task. The feasibility of the proposed model is evaluated using mean accuracy, validation accuracy, and standard deviation. The study results of the very deep neural network model are compared to a convolutional neural network and convolutional neural network with batch normalization. According to the results of the comparison study, very deep neural networks achieve high accuracy of 99.1% for a handwritten dataset. The outcome of the proposed work is used to interpret how well a very deep neural network performs when compared to the other two models of deep neural networks. This proposed architecture may be used to automate the classification of handwritten digits dataset.

Batch Normalization (BatchNorm) is a technique that improves the training of deep neural networks, especially Convolutional Neural Networks (CNN). It has been empirically demonstrated that BatchNorm increases performance, stability, and accuracy, although the reasons for such improvements are unclear. BatchNorm includes a normalization step as well as trainable shift and scale parameters. In this paper, we empirically examine the relative contribution to the success of BatchNorm of the normalization step, as compared to the re-parameterization via shifting and scaling. To conduct our experiments, we implement two new optimizers in PyTorch, namely, a version of BatchNorm that we refer to as AffineLayer, which includes the re-parameterization step without normalization, and a version with just the normalization step, that we call BatchNorm-minus. We compare the performance of our AffineLayer and BatchNorm-minus implementations to standard BatchNorm, and we also compare these to the case where no batch normalization is used. We experiment with four ResNet architectures (ResNet18, ResNet34, ResNet50, and ResNet101) over a standard image dataset and multiple batch sizes. Among other findings, we provide empirical evidence that the success of BatchNorm may derive primarily from improved weight initialization.

The state-of-the-art training of deep neural networks requires to normalize the activities of the neurons for accelerating the training process. A standard approach is to employ batch normalization (BN), in which the activations are normalized by the mean and standard deviation of the training mini-batch. To be invertible, BN also introduces an adaptive gain and bias which are applied after the normalization but often before the non-linearity. In this paper, we investigate the effects of learnable parameters, gain and bias, on the training of various typical deep neural nets, including ALL-CNNs, Network In Network (NIN), ResNets. Through extensive experiments, we show that there is no big difference in both training convergence and final test accuracy if we remove the BN layer following the final convolutional layer from a convolutional neural network (CNN) for standard classification tasks. We also observed that without adaptively updating learnable parameters for BN layers, it often requires less time for training of very deep neural nets such as ResNet-101.