PyTorchKR Papers

최신 논문

총 10,163편
태그·기간 등 상세 필터

카테고리 탐색

로그인 후 이 태그 구독

arXiv:2010.08515

Why Are Convolutional Nets More Sample-Efficient than Fully-Connected Nets?

Convolutional neural networks often dominate fully-connected counterparts in generalization performance, especially on image classification tasks. This is often explained in terms of 'better inductive bias'. However, this has not been made mathematically rigorous, and the hurdle is that the fully connected net can always simulate the convolutional net (for a fixed task). Thus the training algorithm plays a role. The current work describes a natural task on which a provable sample complexity gap can be shown, for standard training algorithms. We construct a single natural distribution on $\mathbb{R}^d\times\{\pm 1\}$ on which any orthogonal-invariant algorithm (i.e. fully-connected networks trained with most gradient-based methods from gaussian initialization) requires $\Omega(d^2)$ samples to generalize while $O(1)$ samples suffice for convolutional architectures. Furthermore, we demonstrate a single target function, learning which on all possible distributions leads to an $O(1)$ vs $\Omega(d^2/\varepsilon)$ gap. The proof relies on the fact that SGD on fully-connected network is orthogonal equivariant. Similar results are achieved for $\ell_2$ regression and adaptive training algorithms, e.g. Adam and AdaGrad, which are only permutation equivariant.

arXiv:2010.08127공식 코드 · Papers with Code

The Deep Bootstrap Framework: Good Online Learners are Good Offline Generalizers

We propose a new framework for reasoning about generalization in deep learning. The core idea is to couple the Real World, where optimizers take stochastic gradient steps on the empirical loss, to an Ideal World, where optimizers take steps on the population loss. This leads to an alternate decomposition of test error into: (1) the Ideal World test error plus (2) the gap between the two worlds. If the gap (2) is universally small, this reduces the problem of generalization in offline learning to the problem of optimization in online learning. We then give empirical evidence that this gap between worlds can be small in realistic deep learning settings, in particular supervised image classification. For example, CNNs generalize better than MLPs on image distributions in the Real World, but this is "because" they optimize faster on the population loss in the Ideal World. This suggests our framework is a useful tool for understanding generalization in deep learning, and lays a foundation for future research in the area.

arXiv:2010.08087

Performance evaluation and application of computation based low-cost homogeneous machine learning model algorithm for image classification

The image classification machine learning model was trained with the intention to predict the category of the input image. While multiple state-of-the-art ensemble model methodologies are openly available, this paper evaluates the performance of a low-cost, simple algorithm that would integrate seamlessly into modern production-grade cloud-based applications. The homogeneous models, trained with the full instead of subsets of data, contains varying hyper-parameters and neural layers from one another. These models' inferences will be processed by the new algorithm, which is loosely based on conditional probability theories. The final output will be evaluated.

arXiv:2010.08031공식 코드 · Papers with Code

QReLU and m-QReLU: Two novel quantum activation functions to aid medical diagnostics

The ReLU activation function (AF) has been extensively applied in deep neural networks, in particular Convolutional Neural Networks (CNN), for image classification despite its unresolved dying ReLU problem, which poses challenges to reliable applications. This issue has obvious important implications for critical applications, such as those in healthcare. Recent approaches are just proposing variations of the activation function within the same unresolved dying ReLU challenge. This contribution reports a different research direction by investigating the development of an innovative quantum approach to the ReLU AF that avoids the dying ReLU problem by disruptive design. The Leaky ReLU was leveraged as a baseline on which the two quantum principles of entanglement and superposition were applied to derive the proposed Quantum ReLU (QReLU) and the modified-QReLU (m-QReLU) activation functions. Both QReLU and m-QReLU are implemented and made freely available in TensorFlow and Keras. This original approach is effective and validated extensively in case studies that facilitate the detection of COVID-19 and Parkinson Disease (PD) from medical images. The two novel AFs were evaluated in a two-layered CNN against nine ReLU-based AFs on seven benchmark datasets, including images of spiral drawings taken via graphic tablets from patients with Parkinson Disease and healthy subjects, and point-of-care ultrasound images on the lungs of patients with COVID-19, those with pneumonia and healthy controls. Despite a higher computational cost, results indicated an overall higher classification accuracy, precision, recall and F1-score brought about by either quantum AFs on five of the seven bench-mark datasets, thus demonstrating its potential to be the new benchmark or gold standard AF in CNNs and aid image classification tasks involved in critical applications, such as medical diagnoses of COVID-19 and PD.

arXiv:2010.06866

Deep Ensembles for Low-Data Transfer Learning

In the low-data regime, it is difficult to train good supervised models from scratch. Instead practitioners turn to pre-trained models, leveraging transfer learning. Ensembling is an empirically and theoretically appealing way to construct powerful predictive models, but the predominant approach of training multiple deep networks with different random initialisations collides with the need for transfer via pre-trained weights. In this work, we study different ways of creating ensembles from pre-trained models. We show that the nature of pre-training itself is a performant source of diversity, and propose a practical algorithm that efficiently identifies a subset of pre-trained models for any downstream dataset. The approach is simple: Use nearest-neighbour accuracy to rank pre-trained models, fine-tune the best ones with a small hyperparameter sweep, and greedily construct an ensemble to minimise validation cross-entropy. When evaluated together with strong baselines on 19 different downstream tasks (the Visual Task Adaptation Benchmark), this achieves state-of-the-art performance at a much lower inference budget, even when selecting from over 2,000 pre-trained models. We also assess our ensembles on ImageNet variants and show improved robustness to distribution shift.

arXiv:2010.06777

Ferrograph image classification

It has been challenging to identify ferrograph images with a small dataset and various scales of wear particle. A novel model is proposed in this study to cope with these challenging problems. For the problem of insufficient samples, we first proposed a data augmentation algorithm based on the permutation of image patches. Then, an auxiliary loss function of image patch permutation recognition was proposed to identify the image generated by the data augmentation algorithm. Moreover, we designed a feature extraction loss function to force the proposed model to extract more abundant features and to reduce redundant representations. As for the challenge of large change range of wear particle size, we proposed a multi-scale feature extraction block to obtain the multi-scale representations of wear particles. We carried out experiments on a ferrograph image dataset and a mini-CIFAR-10 dataset. Experimental results show that the proposed model can improve the accuracy of the two datasets by 9% and 20% respectively compared with the baseline.

arXiv:2010.06682

Are all negatives created equal in contrastive instance discrimination?

Self-supervised learning has recently begun to rival supervised learning on computer vision tasks. Many of the recent approaches have been based on contrastive instance discrimination (CID), in which the network is trained to recognize two augmented versions of the same instance (a query and positive) while discriminating against a pool of other instances (negatives). The learned representation is then used on downstream tasks such as image classification. Using methodology from MoCo v2 (Chen et al., 2020), we divided negatives by their difficulty for a given query and studied which difficulty ranges were most important for learning useful representations. We found a minority of negatives -- the hardest 5% -- were both necessary and sufficient for the downstream task to reach nearly full accuracy. Conversely, the easiest 95% of negatives were unnecessary and insufficient. Moreover, the very hardest 0.1% of negatives were unnecessary and sometimes detrimental. Finally, we studied the properties of negatives that affect their hardness, and found that hard negatives were more semantically similar to the query, and that some negatives were more consistently easy or hard than we would expect by chance. Together, our results indicate that negatives vary in importance and that CID may benefit from more intelligent negative treatment.

arXiv:2010.07384

Human-interpretable model explainability on high-dimensional data

The importance of explainability in machine learning continues to grow, as both neural-network architectures and the data they model become increasingly complex. Unique challenges arise when a model's input features become high dimensional: on one hand, principled model-agnostic approaches to explainability become too computationally expensive; on the other, more efficient explainability algorithms lack natural interpretations for general users. In this work, we introduce a framework for human-interpretable explainability on high-dimensional data, consisting of two modules. First, we apply a semantically meaningful latent representation, both to reduce the raw dimensionality of the data, and to ensure its human interpretability. These latent features can be learnt, e.g. explicitly as disentangled representations or implicitly through image-to-image translation, or they can be based on any computable quantities the user chooses. Second, we adapt the Shapley paradigm for model-agnostic explainability to operate on these latent features. This leads to interpretable model explanations that are both theoretically controlled and computationally tractable. We benchmark our approach on synthetic data and demonstrate its effectiveness on several image-classification tasks.

arXiv:2010.07359공식 코드 · Papers with Code

Effects of the Nonlinearity in Activation Functions on the Performance of Deep Learning Models

The nonlinearity of activation functions used in deep learning models are crucial for the success of predictive models. There are several commonly used simple nonlinear functions, including Rectified Linear Unit (ReLU) and Leaky-ReLU (L-ReLU). In practice, these functions remarkably enhance the model accuracy. However, there is limited insight into the functionality of these nonlinear activation functions in terms of why certain models perform better than others. Here, we investigate the model performance when using ReLU or L-ReLU as activation functions in different model architectures and data domains. Interestingly, we found that the application of L-ReLU is mostly effective when the number of trainable parameters in a model is relatively small. Furthermore, we found that the image classification models seem to perform well with L-ReLU in fully connected layers, especially when pre-trained models such as the VGG-16 are used for the transfer learning.

arXiv:2010.07230공식 코드 · Papers with Code

An Evasion Attack against Stacked Capsule Autoencoder

Capsule network is a type of neural network that uses the spatial relationship between features to classify images. By capturing the poses and relative positions between features, its ability to recognize affine transformation is improved, and it surpasses traditional convolutional neural networks (CNNs) when handling translation, rotation and scaling. The Stacked Capsule Autoencoder (SCAE) is the state-of-the-art capsule network. The SCAE encodes an image as capsules, each of which contains poses of features and their correlations. The encoded contents are then input into the downstream classifier to predict the categories of the images. Existing research mainly focuses on the security of capsule networks with dynamic routing or EM routing, and little attention has been given to the security and robustness of the SCAE. In this paper, we propose an evasion attack against the SCAE. After a perturbation is generated based on the output of the object capsules in the model, it is added to an image to reduce the contribution of the object capsules related to the original category of the image so that the perturbed image will be misclassified. We evaluate the attack using an image classification experiment, and the experimental results indicate that the attack can achieve high success rates and stealthiness. It confirms that the SCAE has a security vulnerability whereby it is possible to craft adversarial samples without changing the original structure of the image to fool the classifiers. We hope that our work will make the community aware of the threat of this attack and raise the attention given to the SCAE's security.

arXiv:2010.07621공식 코드 · Papers with Code

HS-ResNet: Hierarchical-Split Block on Convolutional Neural Network

This paper addresses representational block named Hierarchical-Split Block, which can be taken as a plug-and-play block to upgrade existing convolutional neural networks, improves model performance significantly in a network. Hierarchical-Split Block contains many hierarchical split and concatenate connections within one single residual block. We find multi-scale features is of great importance for numerous vision tasks. Moreover, Hierarchical-Split block is very flexible and efficient, which provides a large space of potential network architectures for different applications. In this work, we present a common backbone based on Hierarchical-Split block for tasks: image classification, object detection, instance segmentation and semantic image segmentation/parsing. Our approach shows significant improvements over all these core tasks in comparison with the baseline. As shown in Figure1, for image classification, our 50-layers network(HS-ResNet50) achieves 81.28% top-1 accuracy with competitive latency on ImageNet-1k dataset. It also outperforms most state-of-the-art models. The source code and models will be available on: https://github.com/PaddlePaddle/PaddleClas

arXiv:2010.07611공식 코드 · Papers with Code

Layer-adaptive sparsity for the Magnitude-based Pruning

Recent discoveries on neural network pruning reveal that, with a carefully chosen layerwise sparsity, a simple magnitude-based pruning achieves state-of-the-art tradeoff between sparsity and performance. However, without a clear consensus on "how to choose," the layerwise sparsities are mostly selected algorithm-by-algorithm, often resorting to handcrafted heuristics or an extensive hyperparameter search. To fill this gap, we propose a novel importance score for global pruning, coined layer-adaptive magnitude-based pruning (LAMP) score; the score is a rescaled version of weight magnitude that incorporates the model-level $\ell_2$ distortion incurred by pruning, and does not require any hyperparameter tuning or heavy computation. Under various image classification setups, LAMP consistently outperforms popular existing schemes for layerwise sparsity selection. Furthermore, we observe that LAMP continues to outperform baselines even in weight-rewinding setups, while the connectivity-oriented layerwise sparsity (the strongest baseline overall) performs worse than a simple global magnitude-based pruning in this case. Code: https://github.com/jaeho-lee/layer-adaptive-sparsity

arXiv:2010.07810

Does Data Augmentation Benefit from Split BatchNorms

Data augmentation has emerged as a powerful technique for improving the performance of deep neural networks and led to state-of-the-art results in computer vision. However, state-of-the-art data augmentation strongly distorts training images, leading to a disparity between examples seen during training and inference. In this work, we explore a recently proposed training paradigm in order to correct for this disparity: using an auxiliary BatchNorm for the potentially out-of-distribution, strongly augmented images. Our experiments then focus on how to define the BatchNorm parameters that are used at evaluation. To eliminate the train-test disparity, we experiment with using the batch statistics defined by clean training images only, yet surprisingly find that this does not yield improvements in model performance. Instead, we investigate using BatchNorm parameters defined by weak augmentations and find that this method significantly improves the performance of common image classification benchmarks such as CIFAR-10, CIFAR-100, and ImageNet. We then explore a fundamental trade-off between accuracy and robustness coming from using different BatchNorm parameters, providing greater insight into the benefits of data augmentation on model performance.

arXiv:2010.07783

Improved Multi-Source Domain Adaptation by Preservation of Factors

Domain Adaptation (DA) is a highly relevant research topic when it comes to image classification with deep neural networks. Combining multiple source domains in a sophisticated way to optimize a classification model can improve the generalization to a target domain. Here, the difference in data distributions of source and target image datasets plays a major role. In this paper, we describe based on a theory of visual factors how real-world scenes appear in images in general and how recent DA datasets are composed of such. We show that different domains can be described by a set of so called domain factors, whose values are consistent within a domain, but can change across domains. Many DA approaches try to remove all domain factors from the feature representation to be domain invariant. In this paper we show that this can lead to negative transfer since task-informative factors can get lost as well. To address this, we propose Factor-Preserving DA (FP-DA), a method to train a deep adversarial unsupervised DA model, which is able to preserve specific task relevant factors in a multi-domain scenario. We demonstrate on CORe50, a dataset with many domains, how such factors can be identified by standard one-to-one transfer experiments between single domains combined with PCA. By applying FP-DA, we show that the highest average and minimum performance can be achieved.

arXiv:2010.07726

LiteDepthwiseNet: An Extreme Lightweight Network for Hyperspectral Image Classification

Deep learning methods have shown considerable potential for hyperspectral image (HSI) classification, which can achieve high accuracy compared with traditional methods. However, they often need a large number of training samples and have a lot of parameters and high computational overhead. To solve these problems, this paper proposes a new network architecture, LiteDepthwiseNet, for HSI classification. Based on 3D depthwise convolution, LiteDepthwiseNet can decompose standard convolution into depthwise convolution and pointwise convolution, which can achieve high classification performance with minimal parameters. Moreover, we remove the ReLU layer and Batch Normalization layer in the original 3D depthwise convolution, which significantly improves the overfitting phenomenon of the model on small sized datasets. In addition, focal loss is used as the loss function to improve the model's attention on difficult samples and unbalanced data, and its training performance is significantly better than that of cross-entropy loss or balanced cross-entropy loss. Experiment results on three benchmark hyperspectral datasets show that LiteDepthwiseNet achieves state-of-the-art performance with a very small number of parameters and low computational cost.

arXiv:2010.07539공식 코드 · Papers with Code

Self-Supervised Domain Adaptation with Consistency Training

We consider the problem of unsupervised domain adaptation for image classification. To learn target-domain-aware features from the unlabeled data, we create a self-supervised pretext task by augmenting the unlabeled data with a certain type of transformation (specifically, image rotation) and ask the learner to predict the properties of the transformation. However, the obtained feature representation may contain a large amount of irrelevant information with respect to the main task. To provide further guidance, we force the feature representation of the augmented data to be consistent with that of the original data. Intuitively, the consistency introduces additional constraints to representation learning, therefore, the learned representation is more likely to focus on the right information about the main task. Our experimental results validate the proposed method and demonstrate state-of-the-art performance on classical domain adaptation benchmarks. Code is available at https://github.com/Jiaolong/ss-da-consistency.

arXiv:2010.07468공식 코드 · Papers with Code

AdaBelief Optimizer: Adapting Stepsizes by the Belief in Observed Gradients

Most popular optimizers for deep learning can be broadly categorized as adaptive methods (e.g. Adam) and accelerated schemes (e.g. stochastic gradient descent (SGD) with momentum). For many models such as convolutional neural networks (CNNs), adaptive methods typically converge faster but generalize worse compared to SGD; for complex settings such as generative adversarial networks (GANs), adaptive methods are typically the default because of their stability.We propose AdaBelief to simultaneously achieve three goals: fast convergence as in adaptive methods, good generalization as in SGD, and training stability. The intuition for AdaBelief is to adapt the stepsize according to the "belief" in the current gradient direction. Viewing the exponential moving average (EMA) of the noisy gradient as the prediction of the gradient at the next time step, if the observed gradient greatly deviates from the prediction, we distrust the current observation and take a small step; if the observed gradient is close to the prediction, we trust it and take a large step. We validate AdaBelief in extensive experiments, showing that it outperforms other methods with fast convergence and high accuracy on image classification and language modeling. Specifically, on ImageNet, AdaBelief achieves comparable accuracy to SGD. Furthermore, in the training of a GAN on Cifar10, AdaBelief demonstrates high stability and improves the quality of generated samples compared to a well-tuned Adam optimizer. Code is available at https://github.com/juntang-zhuang/Adabelief-Optimizer

arXiv:2010.06454

The DongNiao International Birds 10000 Dataset

DongNiao International Birds 10000 (DIB-10K) is a challenging image dataset which has more than 10 thousand different types of birds. It was created to enable the study of machine learning and also ornithology research. DIB-10K does not own the copyright of these images. It only provides thumbnails of images, in a way similar to ImageNet.

arXiv:2010.06497

Satellite Image Classification with Deep Learning

Satellite imagery is important for many applications including disaster response, law enforcement, and environmental monitoring. These applications require the manual identification of objects and facilities in the imagery. Because the geographic expanses to be covered are great and the analysts available to conduct the searches are few, automation is required. Yet traditional object detection and classification algorithms are too inaccurate and unreliable to solve the problem. Deep learning is a family of machine learning algorithms that have shown promise for the automation of such tasks. It has achieved success in image understanding by means of convolutional neural networks. In this paper we apply them to the problem of object and facility recognition in high-resolution, multi-spectral satellite imagery. We describe a deep learning system for classifying objects and facilities from the IARPA Functional Map of the World (fMoW) dataset into 63 different classes. The system consists of an ensemble of convolutional neural networks and additional neural networks that integrate satellite metadata with image features. It is implemented in Python using the Keras and TensorFlow deep learning libraries and runs on a Linux server with an NVIDIA Titan X graphics card. At the time of writing the system is in 2nd place in the fMoW TopCoder competition. Its total accuracy is 83%, the F1 score is 0.797, and it classifies 15 of the classes with accuracies of 95% or better.

arXiv:2010.06440

RMDL: Recalibrated multi-instance deep learning for whole slide gastric image classification

The whole slide histopathology images (WSIs) play a critical role in gastric cancer diagnosis. However, due to the large scale of WSIs and various sizes of the abnormal area, how to select informative regions and analyze them are quite challenging during the automatic diagnosis process. The multi-instance learning based on the most discriminative instances can be of great benefit for whole slide gastric image diagnosis. In this paper, we design a recalibrated multi-instance deep learning method (RMDL) to address this challenging problem. We first select the discriminative instances, and then utilize these instances to diagnose diseases based on the proposed RMDL approach. The designed RMDL network is capable of capturing instance-wise dependencies and recalibrating instance features according to the importance coefficient learned from the fused features. Furthermore, we build a large whole-slide gastric histopathology image dataset with detailed pixel-level annotations. Experimental results on the constructed gastric dataset demonstrate the significant improvement on the accuracy of our proposed framework compared with other state-of-the-art multi-instance learning methods. Moreover, our method is general and can be extended to other diagnosis tasks of different cancer types based on WSIs.