Mutual calibration training: Training deep neural networks with noisy labels using dual-models

Training deep neural networks may be challenging in real world data. Using models as black-boxes, even with transfer learning, can result in poor generalization or inconclusive results when it comes to small datasets or specific applications. This tutorial covers the basic steps as well as more recent options to improve models, in particular, but not restricted to, supervised learning. It can be particularly useful in datasets that are not as well-prepared as those in challenges, and also under scarce annotation and/or small data. We describe basic procedures as data preparation, optimization and transfer learning, but also recent architectural choices such as use of transformer modules, alternative convolutional layers, activation functions, wide/depth, as well as training procedures including curriculum, contrastive and self-supervised learning.

Seismic data denoising is a critical component of seismic data processing, yet effectively removing erratic noise, characterized by its non-Gaussian distribution and high amplitude, remains a substantial challenge for conventional methods and deep-learning (DL) algorithms. Supervised learning frameworks typically outperform others, but they require pairs of noisy data sets alongside corresponding clean ground truth, which is impractical for real-world seismic data sets. In contrast, unsupervised learning (UL) methods, which do not rely on ground truth during training, often fall short in performance when compared with their supervised or traditional denoising counterparts. Moreover, current unsupervised DL methods fail to address the specific challenges posed by erratic seismic noise adequately. This paper introduces a novel zero-shot unsupervised DL framework designed specifically to mitigate random and erratic noise, with a particular emphasis on blended noise. Drawing inspiration from Noise2Noise (N2N) and data augmentation principles, we develop a robust self-supervised denoising network called robust Noiser2Noiser. Our approach eliminates the need for paired noisy and clean data sets as required by supervised methods or paired noisy data sets as in N2N. Instead, our framework relies solely on the original noisy seismic data set. Our methodology generates two independent recorrupted data sets from the original noisy data set, using one as the input and the other as the training target. Subsequently, we use a DL-based denoiser, denoising convolutional neural network, for training purposes. To address various types of random and erratic noise, the original noisy data set is recorrupted with the same noise type. Detailed explanations for generating training input and target data for blended data are provided. We apply our network to synthetic and real marine data examples, demonstrating significantly improved noise attenuation performance compared with traditional denoising methods and state-of-the-art UL Codes are available on https://github.com/Ji-seismic/N2N_deblending.

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.

Purpose: The aim of the article is to develop an algorithm for classifying cracks in the analyzed images using modern methods of deep machine learning and transfer learning based on pretrained convolutional neural network - Inception-ResNet-v2. Design/Methodology/Approach: Transfer learning based on the pretrained convolutional neural network was used to categorize the images. The fully conected layer of the Inception-ResNet-v2 network has been modified. The last layer was trained using a two-class (binary) linear SVM (Support Vector Machine). In total, 20,000 training cases (images) were used to train the fully connected layer within transfer learning process. The research analyzed the possibility of using the deep neural networks for quick and fully automatic identification of cracks / defects on the surface of analyzed parts. Findings: The results indicate that pretrained convolutional neural network using SVM to train a fully connected layer is a very effective solution for visual crack / fault detection. In the analyzed model, a positive classification was obtained at the level of 99.89%. Practical Implications: The model presented in the article can be used in quality control carried out by process monitoring. An effective model for identifying defective parts can be used in both logistics and production processes. Originality/Value: A novelty is the use of a freely available, deep neural network trained to classify 1000 categories of various images for binary categorization of faults (cracks). The algorithm was adjusted by replacing the primary, 1000-output fully connected layer in the Inception-ResNet-v2 network with a binary layer (2 categories). The fully connected layer has been trained using the classification version of the popular SVM learner, but thanks to the combination of this layer with the sophisticated fearure extraction ability of the pre-trained Inception-ResNet-v2 deep network, the resulting predictive model enables the classification of defects with a very high level of accuracy.

We study the eigenvalue distribution of random matrices pertinent to the analysis of deep neural networks. The matrices resemble the product of the sample covariance matrices, however, an important difference is that the analog of the population covariance matrix is now a function of random data matrices (synaptic weight matrices in the deep neural network terminology). The problem has been treated in recent work [J. Pennington, S. Schoenholz and S. Ganguli, The emergence of spectral universality in deep networks, Proc. Mach. Learn. Res. 84 (2018) 1924–1932, arXiv:1802.09979] by using the techniques of free probability theory. Since, however, free probability theory deals with population covariance matrices which are independent of the data matrices, its applicability in this case has to be justified. The justification has been given in [L. Pastur, On random matrices arising in deep neural networks: Gaussian case, Pure Appl. Funct. Anal. (2020), in press, arXiv:2001.06188] for Gaussian data matrices with independent entries, a standard analytical model of free probability, by using a version of the techniques of random matrix theory. In this paper, we use another version of the techniques to extend the results of [L. Pastur, On random matrices arising in deep neural networks: Gaussian case, Pure Appl. Funct. Anal. (2020), in press, arXiv:2001.06188] to the case where the entries of the data matrices are just independent identically distributed random variables with zero mean and finite fourth moment. This, in particular, justifies the mean field approximation in the infinite width limit for the deep untrained neural networks and the property of the macroscopic universality of random matrix theory in this case.

Medical image segmentation tasks hitherto have achieved excellent progresses with large-scale datasets, which empowers us to train potent deep convolutional neural networks (DCNNs). However, labeling such large-scale datasets is laborious and error-prone, which leads the noisy (or incorrect) labels to be an ubiquitous problem in the real-world scenarios. In addition, data collected from different sites usually exhibit significant data distribution shift (or domain shift). As a result, noisy label and domain shift become two common problems in medical imaging application scenarios, especially in medical image segmentation, which degrade the performance of deep learning models significantly. In this paper, we identify a novel problem hidden in medical image segmentation, which is unsupervised domain adaptation on noisy labeled data, and propose a novel algorithm named "Self-Cleansing Unsupervised Domain Adaptation" (S-CDUA) to address such issue. S-CUDA sets up a realistic scenario to solve the above problems simultaneously where training data (i.e., source domain) not only shows domain shift w.r.t. unsupervised test data (i.e., target domain) but also contains noisy labels. The key idea of S-CUDA is to learn noise-excluding and domain invariant knowledge from noisy supervised data, which will be applied on the highly corrupted data for label cleansing and further data-recycling, as well as on the test data with domain shift for supervised propagation. To this end, we propose a novel framework leveraging noisy-label learning and domain adaptation techniques to cleanse the noisy labels and learn from trustable clean samples, thus enabling robust adaptation and prediction on the target domain. Specifically, we train two peer adversarial networks to identify high-confidence clean data and exchange them in companions to eliminate the error accumulation problem and narrow the domain gap simultaneously. In the meantime, the high-confidence noisy data are detected and cleansed in order to reuse the contaminated training data. Therefore, our proposed method can not only cleanse the noisy labels in the training set but also take full advantage of the existing noisy data to update the parameters of the network. For evaluation, we conduct experiments on two popular datasets (REFUGE and Drishti-GS) for optic disc (OD) and optic cup (OC) segmentation, and on another public multi-vendor dataset for spinal cord gray matter (SCGM) segmentation. Experimental results show that our proposed method can cleanse noisy labels efficiently and obtain a model with better generalization performance at the same time, which outperforms previous state-of-the-art methods by large margin. Our code can be found at https://github.com/zzdxjtu/S-cuda.

Nowadays, it has become well known that efficient training of deep neural networks plays a vital role in various successful applications. To achieve this goal, it is impractical to use only one computer, especially when the scale of models is large and some efficient computing resources are available. In this paper, we present a distributed parallel computing framework for training deep belief networks (DBNs) by employing the great power of high-performance clusters (i.e., a system consists of many computers). Motivated by the greedy layer-wise learning algorithm of DBNs, the whole training process is divided layer by layer and distributed to different machines. At the same time, rough representations are exploited to parallelize the training process. By conducting experiments on several large-scale real datasets, the novel algorithms are shown to significantly accelerate the training speed of DBNs while achieving better or competitive prediction accuracy in comparison with the original algorithm.

Deep networks have strong capacities of embedding data into latent representations and finishing following tasks. However, the capacities largely come from high-quality annotated labels, which are expensive to collect. Noisy labels are more affordable, but result in corrupted representations, leading to poor generalization performance. To learn robust representations and handle noisy labels, we propose selective-supervised contrastive learning (Sel-CL) in this paper. Specifically, Sel-CL extend supervised contrastive learning (Sup-CL), which is powerful in representation learning, but is degraded when there are noisy labels. Sel-CL tackles the direct cause of the problem of Sup-CL. That is, as Sup-CL works in a pair-wise manner, noisy pairs built by noisy labels mislead representation learning. To alleviate the issue, we select confident pairs out of noisy ones for Sup-CL without knowing noise rates. In the selection process, by measuring the agreement between learned representations and given labels, we first identify confident examples that are exploited to build confident pairs. Then, the representation similarity distribution in the built confident pairs is exploited to identify more confident pairs out of noisy pairs. All obtained confident pairs are finally used for Sup-CL to enhance representations. Experiments on multiple noisy datasets demonstrate the robustness of the learned representations by our method, following the state-of-the-art performance. Source codes are available at https://github.com/ShikunLi/Sel-Cl.

Learning to rank has recently emerged as an attractive technique to train deep convolutional neural networks for various computer vision tasks. Pairwise ranking, in particular, has been successful in multi-label image classification, achieving state-of-the-art results on various benchmarks. However, most existing approaches use the hinge loss to train their models, which is non-smooth and thus is difficult to optimize especially with deep networks. Furthermore, they employ simple heuristics, such as top-k or thresholding, to determine which labels to include in the output from a ranked list of labels, which limits their use in the real-world setting. In this work, we propose two techniques to improve pairwise ranking based multi-label image classification: (1) we propose a novel loss function for pairwise ranking, which is smooth everywhere and thus is easier to optimize; and (2) we incorporate a label decision module into the model, estimating the optimal confidence thresholds for each visual concept. We provide theoretical analyses of our loss function in the Bayes consistency and risk minimization framework, and show its benefit over existing pairwise ranking formulations. We demonstrate the effectiveness of our approach on three large-scale datasets, VOC2007, NUS-WIDE and MS-COCO, achieving the best reported results in the literature.

The performance of deep neural networks (DNNs) critically relies on high-quality annotations, while training DNNs with noisy labels remains challenging owing to their incredible capacity to memorize the entire training set. In this work, we use two synchronously trained networks to reveal that noisy labels may result in more divergent gradients when updating the parameters. To overcome this, we propose a novel co-training framework named gradient agreement learning (GAL). By dynamically evaluating the gradient agreement coefficient of every pair of parameters from two identical DNNs to determine whether to update them in the training process. GAL can effectively hinder the memorization of noisy labels. Furthermore, we utilize the pseudo labels produced by the two DNNs as the supervision for the training of another network, thereby gaining further improvement by correcting some noisy labels while overcoming the confirmation bias. Extensive experiments on various benchmark datasets demonstrate the superiority of the proposed GAL.

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.

This paper proposes to leverage multiple facets of person photos to improve the training of deep neural networks. Existing studies usually require a lot of labeled images to train deep convolutional networks. Our study suggests exploring multiple datasets and learning effective representation to learn related visual concepts. The practice of learning from multiple facets implicitly enforces to share features for image recognition. We show deep neural network benefits from the learning of multiple person-related categories in photos. Faceted classification systems learn from multiple resources, and alleviate the overfitting problems in deep learning. Moreover, by exploring multiple taxonomies of an object, it provides a finer annotation for the query images.

In recent years, deep neural networks have been successfully applied to model visual concepts and have achieved competitive performance on many tasks. Despite their impressive performance, traditional deep networks are subjected to the decayed performance under the condition of lacking sufficient training data. This problem becomes extremely severe for deep networks trained on a very small dataset, making them overfitting by capturing nonessential or noisy information in the training set. Toward this end, we propose a novel generalized deep transfer networks (DTNs), capable of transferring label information across heterogeneous domains, textual domain to visual domain. The proposed framework has the ability to adequately mitigate the problem of insufficient training images by bringing in rich labels from the textual domain. Specifically, to share the labels between two domains, we build parameter- and representation-shared layers. They are able to generate domain-specific and shared interdomain features, making this architecture flexible and powerful in capturing complex information from different domains jointly. To evaluate the proposed method, we release a new dataset extended from NUS-WIDE at http://imag.njust.edu.cn/NUS-WIDE-128.html. Experimental results on this dataset show the superior performance of the proposed DTNs compared to existing state-of-the-art methods.

Layer-by-layer training is an alternative end-to-end backpropagation approach for training deep convolutional neural networks. Layer-by-layer training on specific architectures can yield very competitive results. In the ImageNet database (www.imagenet. org), layer-by-layer trained networks can perform comparable to many modern end-to-end trained networks. This article compares the performance gap between the two training procedures over a wide range of network architectures and further analyzes the potential limitations of layer-by-layer training. The results show that layer-by-layer learning quickly saturates after a certain critical level due to overfitting of early levels in neural networks. Several approaches that have been used to solve this problem are discussed, and a methodology for improving layer-by-layer learning in various neural network architectures is discussed. Fundamentally, this research highlights the need to open up the black box that represents modern deep neural networks and explore the layer-by-layer interactions between intermediate hidden layers within deep networks through the lens of layer-by-layer learning.

Training deep neural network (DNN) with noisy labels is practically challenging since inaccurate labels severely degrade the generalization ability of DNN. Previous efforts tend to handle part or full data in a unified denoising flow via identifying noisy data with a coarse small-loss criterion to mitigate the interference from noisy labels, ignoring the fact that the difficulties of noisy samples are different, thus a rigid and unified data selection pipeline cannot tackle this problem well . In this paper, we first propose a coarse-to-fine robust learning method called CREMA, to handle noisy data in a divide-and-conquer manner. In coarse-level, clean and noisy sets are firstly separated in terms of credibility in a statistical sense. Since it is practically impossible to categorize all noisy samples correctly, we further process them in a fine-grained manner via modeling the credibility of each sample. Specifically, for the clean set, we deliberately design a memory-based modulation scheme to dynamically adjust the contribution of each sample in terms of its historical credibility sequence during training, thus alleviating the effect from noisy samples incorrectly grouped into the clean set. Meanwhile, for samples categorized into the noisy set, a selective label update strategy is proposed to correct noisy labels while mitigating the problem of correction error. Extensive experiments are conducted on benchmarks of different modality, including image classification (CIFAR, Clothing1M etc.) and text recognition (IMDB), with either synthetic or natural semantic noises, demonstrating the superiority and generality of CREMA.