PyTorchKR Papers

최신 논문

총 10,163편
태그·기간 등 상세 필터

카테고리 탐색

로그인 후 이 태그 구독

arXiv:2004.11783

Quantization of Deep Neural Networks for Accumulator-constrained Processors

We introduce an Artificial Neural Network (ANN) quantization methodology for platforms without wide accumulation registers. This enables fixed-point model deployment on embedded compute platforms that are not specifically designed for large kernel computations (i.e. accumulator-constrained processors). We formulate the quantization problem as a function of accumulator size, and aim to maximize the model accuracy by maximizing bit width of input data and weights. To reduce the number of configurations to consider, only solutions that fully utilize the available accumulator bits are being tested. We demonstrate that 16-bit accumulators are able to obtain a classification accuracy within 1\% of the floating-point baselines on the CIFAR-10 and ILSVRC2012 image classification benchmarks. Additionally, a near-optimal $2\times$ speedup is obtained on an ARM processor, by exploiting 16-bit accumulators for image classification on the All-CNN-C and AlexNet networks.

arXiv:2004.11336

Self-supervised Learning for Astronomical Image Classification

In Astronomy, a huge amount of image data is generated daily by photometric surveys, which scan the sky to collect data from stars, galaxies and other celestial objects. In this paper, we propose a technique to leverage unlabeled astronomical images to pre-train deep convolutional neural networks, in order to learn a domain-specific feature extractor which improves the results of machine learning techniques in setups with small amounts of labeled data available. We show that our technique produces results which are in many cases better than using ImageNet pre-training.

arXiv:2004.11362공식 코드 · Papers with Code

Supervised Contrastive Learning

Contrastive learning applied to self-supervised representation learning has seen a resurgence in recent years, leading to state of the art performance in the unsupervised training of deep image models. Modern batch contrastive approaches subsume or significantly outperform traditional contrastive losses such as triplet, max-margin and the N-pairs loss. In this work, we extend the self-supervised batch contrastive approach to the fully-supervised setting, allowing us to effectively leverage label information. Clusters of points belonging to the same class are pulled together in embedding space, while simultaneously pushing apart clusters of samples from different classes. We analyze two possible versions of the supervised contrastive (SupCon) loss, identifying the best-performing formulation of the loss. On ResNet-200, we achieve top-1 accuracy of 81.4% on the ImageNet dataset, which is 0.8% above the best number reported for this architecture. We show consistent outperformance over cross-entropy on other datasets and two ResNet variants. The loss shows benefits for robustness to natural corruptions and is more stable to hyperparameter settings such as optimizers and data augmentations. Our loss function is simple to implement, and reference TensorFlow code is released at https://t.ly/supcon.

arXiv:2004.11149

A Comprehensive Overview and Survey of Recent Advances in Meta-Learning

This article reviews meta-learning also known as learning-to-learn which seeks rapid and accurate model adaptation to unseen tasks with applications in highly automated AI, few-shot learning, natural language processing and robotics. Unlike deep learning, meta-learning can be applied to few-shot high-dimensional datasets and considers further improving model generalization to unseen tasks. Deep learning is focused upon in-sample prediction and meta-learning concerns model adaptation for out-of-sample prediction. Meta-learning can continually perform self-improvement to achieve highly autonomous AI. Meta-learning may serve as an additional generalization block complementary for original deep learning model. Meta-learning seeks adaptation of machine learning models to unseen tasks which are vastly different from trained tasks. Meta-learning with coevolution between agent and environment provides solutions for complex tasks unsolvable by training from scratch. Meta-learning methodology covers a wide range of great minds and thoughts. We briefly introduce meta-learning methodologies in the following categories: black-box meta-learning, metric-based meta-learning, layered meta-learning and Bayesian meta-learning framework. Recent applications concentrate upon the integration of meta-learning with other machine learning framework to provide feasible integrated problem solutions. We briefly present recent meta-learning advances and discuss potential future research directions.

arXiv:2004.11116

Deep Learning Classification With Noisy Labels

Deep Learning systems have shown tremendous accuracy in image classification, at the cost of big image datasets. Collecting such amounts of data can lead to labelling errors in the training set. Indexing multimedia content for retrieval, classification or recommendation can involve tagging or classification based on multiple criteria. In our case, we train face recognition systems for actors identification with a closed set of identities while being exposed to a significant number of perturbators (actors unknown to our database). Face classifiers are known to be sensitive to label noise. We review recent works on how to manage noisy annotations when training deep learning classifiers, independently from our interest in face recognition.

arXiv:2004.10968

ArchNet: Data Hiding Model in Distributed Machine Learning System

Integrating idle embedded devices into cloud computing is a promising approach to support distributed machine learning. In this paper, we approach to address the data hiding problem in such distributed machine learning systems. For the purpose of the data encryption in the distributed machine learning systems, we propose the Tripartite Asymmetric Encryption theorem and give mathematical proof. Based on the theorem, we design a general image encryption scheme ArchNet.The scheme has been implemented on MNIST, Fashion-MNIST and Cifar-10 datasets to simulate real situation. We use different base models on the encrypted datasets and compare the results with the RC4 algorithm and differential privacy policy. Experiment results evaluated the efficiency of the proposed design. Specifically, our design can improve the accuracy on MNIST up to 97.26% compared with RC4.The accuracies on the datasets encrypted by ArchNet are 97.26%, 84.15% and 79.80%, and they are 97.31%, 82.31% and 80.22% on the original datasets, which shows that the encrypted accuracy of ArchNet has the same performance as the base model. It also shows that ArchNet can be deployed on the distributed system with embedded devices.

arXiv:2004.10780

Diagram Image Retrieval using Sketch-Based Deep Learning and Transfer Learning

Resolution of the complex problem of image retrieval for diagram images has yet to be reached. Deep learning methods continue to excel in the fields of object detection and image classification applied to natural imagery. However, the application of such methodologies applied to binary imagery remains limited due to lack of crucial features such as textures,color and intensity information. This paper presents a deep learning based method for image-based search for binary patent images by taking advantage of existing large natural image repositories for image search and sketch-based methods (Sketches are not identical to diagrams, but they do share some characteristics; for example, both imagery types are gray scale (binary), composed of contours, and are lacking in texture). We begin by using deep learning to generate sketches from natural images for image retrieval and then train a second deep learning model on the sketches. We then use our small set of manually labeled patent diagram images via transfer learning to adapt the image search from sketches of natural images to diagrams. Our experiment results show the effectiveness of deep learning with transfer learning for detecting near-identical copies in patent images and querying similar images based on content.

arXiv:2004.10705

A Committee of Convolutional Neural Networks for Image Classication in the Concurrent Presence of Feature and Label Noise

Image classification has become a ubiquitous task. Models trained on good quality data achieve accuracy which in some application domains is already above human-level performance. Unfortunately, real-world data are quite often degenerated by the noise existing in features and/or labels. There are quite many papers that handle the problem of either feature or label noise, separately. However, to the best of our knowledge, this piece of research is the first attempt to address the problem of concurrent occurrence of both types of noise. Basing on the MNIST, CIFAR-10 and CIFAR-100 datasets, we experimentally proved that the difference by which committees beat single models increases along with noise level, no matter it is an attribute or label disruption. Thus, it makes ensembles legitimate to be applied to noisy images with noisy labels. The aforementioned committees' advantage over single models is positively correlated with dataset difficulty level as well. We propose three committee selection algorithms that outperform a strong baseline algorithm which relies on an ensemble of individual (nonassociated) best models.

arXiv:2004.10703공식 코드 · Papers with Code

ktrain: A Low-Code Library for Augmented Machine Learning

We present ktrain, a low-code Python library that makes machine learning more accessible and easier to apply. As a wrapper to TensorFlow and many other libraries (e.g., transformers, scikit-learn, stellargraph), it is designed to make sophisticated, state-of-the-art machine learning models simple to build, train, inspect, and apply by both beginners and experienced practitioners. Featuring modules that support text data (e.g., text classification, sequence tagging, open-domain question-answering), vision data (e.g., image classification), graph data (e.g., node classification, link prediction), and tabular data, ktrain presents a simple unified interface enabling one to quickly solve a wide range of tasks in as little as three or four "commands" or lines of code.

arXiv:2004.10605

Self-Supervised Representation Learning on Document Images

This work analyses the impact of self-supervised pre-training on document images in the context of document image classification. While previous approaches explore the effect of self-supervision on natural images, we show that patch-based pre-training performs poorly on document images because of their different structural properties and poor intra-sample semantic information. We propose two context-aware alternatives to improve performance on the Tobacco-3482 image classification task. We also propose a novel method for self-supervision, which makes use of the inherent multi-modality of documents (image and text), which performs better than other popular self-supervised methods, including supervised ImageNet pre-training, on document image classification scenarios with a limited amount of data.

arXiv:2004.10484공식 코드 · Papers with Code

Understanding Integrated Gradients with SmoothTaylor for Deep Neural Network Attribution

Integrated Gradients as an attribution method for deep neural network models offers simple implementability. However, it suffers from noisiness of explanations which affects the ease of interpretability. The SmoothGrad technique is proposed to solve the noisiness issue and smoothen the attribution maps of any gradient-based attribution method. In this paper, we present SmoothTaylor as a novel theoretical concept bridging Integrated Gradients and SmoothGrad, from the Taylor's theorem perspective. We apply the methods to the image classification problem, using the ILSVRC2012 ImageNet object recognition dataset, and a couple of pretrained image models to generate attribution maps. These attribution maps are empirically evaluated using quantitative measures for sensitivity and noise level. We further propose adaptive noising to optimize for the noise scale hyperparameter value. From our experiments, we find that the SmoothTaylor approach together with adaptive noising is able to generate better quality saliency maps with lesser noise and higher sensitivity to the relevant points in the input space as compared to Integrated Gradients.

arXiv:2004.10664

A review: Deep learning for medical image segmentation using multi-modality fusion

Multi-modality is widely used in medical imaging, because it can provide multiinformation about a target (tumor, organ or tissue). Segmentation using multimodality consists of fusing multi-information to improve the segmentation. Recently, deep learning-based approaches have presented the state-of-the-art performance in image classification, segmentation, object detection and tracking tasks. Due to their self-learning and generalization ability over large amounts of data, deep learning recently has also gained great interest in multi-modal medical image segmentation. In this paper, we give an overview of deep learning-based approaches for multi-modal medical image segmentation task. Firstly, we introduce the general principle of deep learning and multi-modal medical image segmentation. Secondly, we present different deep learning network architectures, then analyze their fusion strategies and compare their results. The earlier fusion is commonly used, since it's simple and it focuses on the subsequent segmentation network architecture. However, the later fusion gives more attention on fusion strategy to learn the complex relationship between different modalities. In general, compared to the earlier fusion, the later fusion can give more accurate result if the fusion method is effective enough. We also discuss some common problems in medical image segmentation. Finally, we summarize and provide some perspectives on the future research.

arXiv:2004.09694

Alleviating the Incompatibility between Cross Entropy Loss and Episode Training for Few-shot Skin Disease Classification

Skin disease classification from images is crucial to dermatological diagnosis. However, identifying skin lesions involves a variety of aspects in terms of size, color, shape, and texture. To make matters worse, many categories only contain very few samples, posing great challenges to conventional machine learning algorithms and even human experts. Inspired by the recent success of Few-Shot Learning (FSL) in natural image classification, we propose to apply FSL to skin disease identification to address the extreme scarcity of training sample problem. However, directly applying FSL to this task does not work well in practice, and we find that the problem can be largely attributed to the incompatibility between Cross Entropy (CE) and episode training, which are both commonly used in FSL. Based on a detailed analysis, we propose the Query-Relative (QR) loss, which proves superior to CE under episode training and is closely related to recently proposed mutual information estimation. Moreover, we further strengthen the proposed QR loss with a novel adaptive hard margin strategy. Comprehensive experiments validate the effectiveness of the proposed FSL scheme and the possibility to diagnosis rare skin disease with a few labeled samples.

arXiv:2004.09576

LSQ+: Improving low-bit quantization through learnable offsets and better initialization

Unlike ReLU, newer activation functions (like Swish, H-swish, Mish) that are frequently employed in popular efficient architectures can also result in negative activation values, with skewed positive and negative ranges. Typical learnable quantization schemes [PACT, LSQ] assume unsigned quantization for activations and quantize all negative activations to zero which leads to significant loss in performance. Naively using signed quantization to accommodate these negative values requires an extra sign bit which is expensive for low-bit (2-, 3-, 4-bit) quantization. To solve this problem, we propose LSQ+, a natural extension of LSQ, wherein we introduce a general asymmetric quantization scheme with trainable scale and offset parameters that can learn to accommodate the negative activations. Gradient-based learnable quantization schemes also commonly suffer from high instability or variance in the final training performance, hence requiring a great deal of hyper-parameter tuning to reach a satisfactory performance. LSQ+ alleviates this problem by using an MSE-based initialization scheme for the quantization parameters. We show that this initialization leads to significantly lower variance in final performance across multiple training runs. Overall, LSQ+ shows state-of-the-art results for EfficientNet and MixNet and also significantly outperforms LSQ for low-bit quantization of neural nets with Swish activations (e.g.: 1.8% gain with W4A4 quantization and upto 5.6% gain with W2A2 quantization of EfficientNet-B0 on ImageNet dataset). To the best of our knowledge, ours is the first work to quantize such architectures to extremely low bit-widths.

arXiv:2004.09506

Revisiting Initialization of Neural Networks

The proper initialization of weights is crucial for the effective training and fast convergence of deep neural networks (DNNs). Prior work in this area has mostly focused on balancing the variance among weights per layer to maintain stability of (i) the input data propagated forwards through the network and (ii) the loss gradients propagated backwards, respectively. This prevalent heuristic is however agnostic of dependencies among gradients across the various layers and captures only firstorder effects. In this paper, we propose and discuss an initialization principle that is based on a rigorous estimation of the global curvature of weights across layers by approximating and controlling the norm of their Hessian matrix. The proposed approach is more systematic and recovers previous results for DNN activations such as smooth functions, dropouts, and ReLU. Our experiments on Word2Vec and the MNIST/CIFAR image classification tasks confirm that tracking the Hessian norm is a useful diagnostic tool which helps to more rigorously initialize weights

arXiv:2004.10076공식 코드 · Papers with Code

Tensor Networks for Medical Image Classification

With the increasing adoption of machine learning tools like neural networks across several domains, interesting connections and comparisons to concepts from other domains are coming to light. In this work, we focus on the class of Tensor Networks, which has been a work horse for physicists in the last two decades to analyse quantum many-body systems. Building on the recent interest in tensor networks for machine learning, we extend the Matrix Product State tensor networks (which can be interpreted as linear classifiers operating in exponentially high dimensional spaces) to be useful in medical image analysis tasks. We focus on classification problems as a first step where we motivate the use of tensor networks and propose adaptions for 2D images using classical image domain concepts such as local orderlessness of images. With the proposed locally orderless tensor network model (LoTeNet), we show that tensor networks are capable of attaining performance that is comparable to state-of-the-art deep learning methods. We evaluate the model on two publicly available medical imaging datasets and show performance improvements with fewer model hyperparameters and lesser computational resources compared to relevant baseline methods.

arXiv:2004.09817

Lottery Hypothesis based Unsupervised Pre-training for Model Compression in Federated Learning

Federated learning (FL) enables a neural network (NN) to be trained using privacy-sensitive data on mobile devices while retaining all the data on their local storages. However, FL asks the mobile devices to perform heavy communication and computation tasks, i.e., devices are requested to upload and download large-volume NN models and train them. This paper proposes a novel unsupervised pre-training method adapted for FL, which aims to reduce both the communication and computation costs through model compression. Since the communication and computation costs are highly dependent on the volume of NN models, reducing the volume without decreasing model performance can reduce these costs. The proposed pre-training method leverages unlabeled data, which is expected to be obtained from the Internet or data repository much more easily than labeled data. The key idea of the proposed method is to obtain a ``good'' subnetwork from the original NN using the unlabeled data based on the lottery hypothesis. The proposed method trains an original model using a denoising auto encoder with the unlabeled data and then prunes small-magnitude parameters of the original model to generate a small but good subnetwork. The proposed method is evaluated using an image classification task. The results show that the proposed method requires 35\% less traffic and computation time than previous methods when achieving a certain test accuracy.

arXiv:2004.09805공식 코드 · Papers with Code

AMC-Loss: Angular Margin Contrastive Loss for Improved Explainability in Image Classification

Deep-learning architectures for classification problems involve the cross-entropy loss sometimes assisted with auxiliary loss functions like center loss, contrastive loss and triplet loss. These auxiliary loss functions facilitate better discrimination between the different classes of interest. However, recent studies hint at the fact that these loss functions do not take into account the intrinsic angular distribution exhibited by the low-level and high-level feature representations. This results in less compactness between samples from the same class and unclear boundary separations between data clusters of different classes. In this paper, we address this issue by proposing the use of geometric constraints, rooted in Riemannian geometry. Specifically, we propose Angular Margin Contrastive Loss (AMC-Loss), a new loss function to be used along with the traditional cross-entropy loss. The AMC-Loss employs the discriminative angular distance metric that is equivalent to geodesic distance on a hypersphere manifold such that it can serve a clear geometric interpretation. We demonstrate the effectiveness of AMC-Loss by providing quantitative and qualitative results. We find that although the proposed geometrically constrained loss-function improves quantitative results modestly, it has a qualitatively surprisingly beneficial effect on increasing the interpretability of deep-net decisions as seen by the visual explanations generated by techniques such as the Grad-CAM. Our code is available at https://github.com/hchoi71/AMC-Loss.

arXiv:2004.08514공식 코드 · Papers with Code

DMT: Dynamic Mutual Training for Semi-Supervised Learning

Recent semi-supervised learning methods use pseudo supervision as core idea, especially self-training methods that generate pseudo labels. However, pseudo labels are unreliable. Self-training methods usually rely on single model prediction confidence to filter low-confidence pseudo labels, thus remaining high-confidence errors and wasting many low-confidence correct labels. In this paper, we point out it is difficult for a model to counter its own errors. Instead, leveraging inter-model disagreement between different models is a key to locate pseudo label errors. With this new viewpoint, we propose mutual training between two different models by a dynamically re-weighted loss function, called Dynamic Mutual Training (DMT). We quantify inter-model disagreement by comparing predictions from two different models to dynamically re-weight loss in training, where a larger disagreement indicates a possible error and corresponds to a lower loss value. Extensive experiments show that DMT achieves state-of-the-art performance in both image classification and semantic segmentation. Our codes are released at https://github.com/voldemortX/DST-CBC .

arXiv:2004.09034

Learning What Makes a Difference from Counterfactual Examples and Gradient Supervision

One of the primary challenges limiting the applicability of deep learning is its susceptibility to learning spurious correlations rather than the underlying mechanisms of the task of interest. The resulting failure to generalise cannot be addressed by simply using more data from the same distribution. We propose an auxiliary training objective that improves the generalization capabilities of neural networks by leveraging an overlooked supervisory signal found in existing datasets. We use pairs of minimally-different examples with different labels, a.k.a counterfactual or contrasting examples, which provide a signal indicative of the underlying causal structure of the task. We show that such pairs can be identified in a number of existing datasets in computer vision (visual question answering, multi-label image classification) and natural language processing (sentiment analysis, natural language inference). The new training objective orients the gradient of a model's decision function with pairs of counterfactual examples. Models trained with this technique demonstrate improved performance on out-of-distribution test sets.