PyTorchKR Papers

최신 논문

총 10,163편
태그·기간 등 상세 필터

카테고리 탐색

로그인 후 이 태그 구독

arXiv:1911.09659

AdaFilter: Adaptive Filter Fine-tuning for Deep Transfer Learning

There is an increasing number of pre-trained deep neural network models. However, it is still unclear how to effectively use these models for a new task. Transfer learning, which aims to transfer knowledge from source tasks to a target task, is an effective solution to this problem. Fine-tuning is a popular transfer learning technique for deep neural networks where a few rounds of training are applied to the parameters of a pre-trained model to adapt them to a new task. Despite its popularity, in this paper, we show that fine-tuning suffers from several drawbacks. We propose an adaptive fine-tuning approach, called AdaFilter, which selects only a part of the convolutional filters in the pre-trained model to optimize on a per-example basis. We use a recurrent gated network to selectively fine-tune convolutional filters based on the activations of the previous layer. We experiment with 7 public image classification datasets and the results show that AdaFilter can reduce the average classification error of the standard fine-tuning by 2.54%.

arXiv:1911.09464공식 코드 · Papers with Code

Quantization Networks

Although deep neural networks are highly effective, their high computational and memory costs severely challenge their applications on portable devices. As a consequence, low-bit quantization, which converts a full-precision neural network into a low-bitwidth integer version, has been an active and promising research topic. Existing methods formulate the low-bit quantization of networks as an approximation or optimization problem. Approximation-based methods confront the gradient mismatch problem, while optimization-based methods are only suitable for quantizing weights and could introduce high computational cost in the training stage. In this paper, we propose a novel perspective of interpreting and implementing neural network quantization by formulating low-bit quantization as a differentiable non-linear function (termed quantization function). The proposed quantization function can be learned in a lossless and end-to-end manner and works for any weights and activations of neural networks in a simple and uniform way. Extensive experiments on image classification and object detection tasks show that our quantization networks outperform the state-of-the-art methods. We believe that the proposed method will shed new insights on the interpretation of neural network quantization. Our code is available at https://github.com/aliyun/alibabacloud-quantization-networks.

arXiv:1911.09418

MSD: Multi-Self-Distillation Learning via Multi-classifiers within Deep Neural Networks

As the development of neural networks, more and more deep neural networks are adopted in various tasks, such as image classification. However, as the huge computational overhead, these networks could not be applied on mobile devices or other low latency scenes. To address this dilemma, multi-classifier convolutional network is proposed to allow faster inference via early classifiers with the corresponding classifiers. These networks utilize sophisticated designing to increase the early classifier accuracy. However, naively training the multi-classifier network could hurt the performance (accuracy) of deep neural networks as early classifiers throughout interfere with the feature generation process. In this paper, we propose a general training framework named multi-self-distillation learning (MSD), which mining knowledge of different classifiers within the same network and increase every classifier accuracy. Our approach can be applied not only to multi-classifier networks, but also modern CNNs (e.g., ResNet Series) augmented with additional side branch classifiers. We use sampling-based branch augmentation technique to transform a single-classifier network into a multi-classifier network. This reduces the gap of capacity between different classifiers, and improves the effectiveness of applying MSD. Our experiments show that MSD improves the accuracy of various networks: enhancing the accuracy of every classifier significantly for existing multi-classifier network (MSDNet), improving vanilla single-classifier networks with internal classifiers with high accuracy, while also improving the final accuracy.

arXiv:1911.08916

Deep Learning based HEp-2 Image Classification: A Comprehensive Review

Classification of HEp-2 cell patterns plays a significant role in the indirect immunofluorescence test for identifying autoimmune diseases in the human body. Many automatic HEp-2 cell classification methods have been proposed in recent years, amongst which deep learning based methods have shown impressive performance. This paper provides a comprehensive review of the existing deep learning based HEp-2 cell image classification methods. These methods perform HEp-2 image classification at two levels, namely, cell-level and specimen-level. Both levels are covered in this review. At each level, the methods are organized with a deep network usage based taxonomy. The core idea, notable achievements, and key strengths and weaknesses of each method are critically analyzed. Furthermore, a concise review of the existing HEp-2 datasets that are commonly used in the literature is given. The paper ends with a discussion on novel opportunities and future research directions in this field. It is hoped that this paper would provide readers with a thorough reference of this novel, challenging, and thriving field.

arXiv:1911.08769

Inspect Transfer Learning Architecture with Dilated Convolution

There are many award-winning pre-trained Convolutional Neural Network (CNN), which have a common phenomenon of increasing depth in convolutional layers. However, I inspect on VGG network, which is one of the famous model submitted to ILSVRC-2014, to show that slight modification in the basic architecture can enhance the accuracy result of the image classification task. In this paper, We present two improve architectures of pre-trained VGG-16 and VGG-19 networks that apply transfer learning when trained on a different dataset. I report a series of experimental result on various modification of the primary VGG networks and achieved significant out-performance on image classification task by: (1) freezing the first two blocks of the convolutional layers to prevent over-fitting and (2) applying different combination of dilation rate in the last three blocks of convolutional layer to reduce image resolution for feature extraction. Both the proposed architecture achieves a competitive result on CIFAR-10 and CIFAR-100 dataset.

arXiv:1911.09032공식 코드 · Papers with Code

Outside the Box: Abstraction-Based Monitoring of Neural Networks

Neural networks have demonstrated unmatched performance in a range of classification tasks. Despite numerous efforts of the research community, novelty detection remains one of the significant limitations of neural networks. The ability to identify previously unseen inputs as novel is crucial for our understanding of the decisions made by neural networks. At runtime, inputs not falling into any of the categories learned during training cannot be classified correctly by the neural network. Existing approaches treat the neural network as a black box and try to detect novel inputs based on the confidence of the output predictions. However, neural networks are not trained to reduce their confidence for novel inputs, which limits the effectiveness of these approaches. We propose a framework to monitor a neural network by observing the hidden layers. We employ a common abstraction from program analysis - boxes - to identify novel behaviors in the monitored layers, i.e., inputs that cause behaviors outside the box. For each neuron, the boxes range over the values seen in training. The framework is efficient and flexible to achieve a desired trade-off between raising false warnings and detecting novel inputs. We illustrate the performance and the robustness to variability in the unknown classes on popular image-classification benchmarks.

arXiv:1911.08953공식 코드 · Papers with Code

MetH: A family of high-resolution and variable-shape image challenges

High-resolution and variable-shape images have not yet been properly addressed by the AI community. The approach of down-sampling data often used with convolutional neural networks is sub-optimal for many tasks, and has too many drawbacks to be considered a sustainable alternative. In sight of the increasing importance of problems that can benefit from exploiting high-resolution (HR) and variable-shape, and with the goal of promoting research in that direction, we introduce a new family of datasets (MetH). The four proposed problems include two image classification, one image regression and one super resolution task. Each of these datasets contains thousands of art pieces captured by HR and variable-shape images, labeled by experts at the Metropolitan Museum of Art. We perform an analysis, which shows how the proposed tasks go well beyond current public alternatives in both pixel size and aspect ratio variance. At the same time, the performance obtained by popular architectures on these tasks shows that there is ample room for improvement. To wrap up the relevance of the contribution we review the fields, both in AI and high-performance computing, that could benefit from the proposed challenges.

arXiv:1911.08907공식 코드 · Papers with Code

Auto-Precision Scaling for Distributed Deep Learning

It has been reported that the communication cost for synchronizing gradients can be a bottleneck, which limits the scalability of distributed deep learning. Using low-precision gradients is a promising technique for reducing the bandwidth requirement. In this work, we propose Auto Precision Scaling (APS), an algorithm that can improve the accuracy when we communicate gradients by low-precision floating-point values. APS can improve the accuracy for all precisions with a trivial communication cost. Our experimental results show that for many applications, APS can train state-of-the-art models by 8-bit gradients with no or only a tiny accuracy loss (<0.05%). Furthermore, we can avoid any accuracy loss by designing a hybrid-precision technique. Finally, we propose a performance model to evaluate the proposed method. Our experimental results show that APS can get a significant speedup over state-of-the-art methods. To make it available to researchers and developers, we design and implement CPD (Customized-Precision Deep Learning) system, which can simulate the training process using an arbitrary low-precision customized floating-point format. We integrate CPD into PyTorch and make it open-source.

arXiv:1911.08264

Visualization approach to assess the robustness of neural networks for medical image classification

The use of neural networks for diagnosis classification is becoming more and more prevalent in the medical imaging community. However, deep learning method outputs remain hard to explain. Another difficulty is to choose among the large number of techniques developed to analyze how networks learn, as all present different limitations. In this paper, we extended the framework of Fong and Vedaldi [IEEE International Conference on Computer Vision (ICCV), 2017] to visualize the training of convolutional neural networks (CNNs) on 3D quantitative neuroimaging data. Our application focuses on the detection of Alzheimer's disease with gray matter probability maps extracted from structural MRI. We first assessed the robustness of the visualization method by studying the coherence of the longitudinal patterns and regions identified by the network. We then studied the stability of the CNN training by computing visualization-based similarity indexes between different re-runs of the CNN. We demonstrated that the areas identified by the CNN were consistent with what is known of Alzheimer's disease and that the visualization approach extract coherent longitudinal patterns. We also showed that the CNN training is not stable and that the areas identified mainly depend on the initialization and the training process. This issue may exist in many other medical studies using deep learning methods on datasets in which the number of samples is too small and the data dimension is high. This means that it may not be possible to rely on deep learning to detect stable regions of interest in this field yet.

arXiv:1911.08252

IC-Network: Efficient Structure for Convolutional Neural Networks

Neural networks have been widely used, and most networks achieve excellent performance by stacking certain types of basic units. Compared to increasing the depth and width of the network, designing more effective basic units has become an important research topic. Inspired by the elastic collision model in physics, we present a universal structure that could be integrated into the existing network structures to speed up the training process and increase their generalization abilities. We term this structure the "Inter-layer Collision" (IC) structure. We built two kinds of basic computational units (IC layer and IC block) that compose the convolutional neural networks (CNNs) by combining the IC structure with the convolution operation. Compared to traditional convolutions, both of the proposed computational units have a stronger non-linear representation ability and can filter features useful for a given task. Using these computational units to build networks, we bring significant improvements in performance for existing state-of-the-art CNNs. On the imagenet experiment, we integrate the IC block into ResNet-50 and reduce the top-1 error from 22.85% to 21.49%, which also exceeds the top-1 error of ResNet-100 (21.75%).

arXiv:1911.08251공식 코드 · Papers with Code

General $E(2)$-Equivariant Steerable CNNs

The big empirical success of group equivariant networks has led in recent years to the sprouting of a great variety of equivariant network architectures. A particular focus has thereby been on rotation and reflection equivariant CNNs for planar images. Here we give a general description of $E(2)$-equivariant convolutions in the framework of Steerable CNNs. The theory of Steerable CNNs thereby yields constraints on the convolution kernels which depend on group representations describing the transformation laws of feature spaces. We show that these constraints for arbitrary group representations can be reduced to constraints under irreducible representations. A general solution of the kernel space constraint is given for arbitrary representations of the Euclidean group $E(2)$ and its subgroups. We implement a wide range of previously proposed and entirely new equivariant network architectures and extensively compare their performances. $E(2)$-steerable convolutions are further shown to yield remarkable gains on CIFAR-10, CIFAR-100 and STL-10 when used as a drop-in replacement for non-equivariant convolutions.

arXiv:1911.08177공식 코드 · Papers with Code

Rethinking deep active learning: Using unlabeled data at model training

Active learning typically focuses on training a model on few labeled examples alone, while unlabeled ones are only used for acquisition. In this work we depart from this setting by using both labeled and unlabeled data during model training across active learning cycles. We do so by using unsupervised feature learning at the beginning of the active learning pipeline and semi-supervised learning at every active learning cycle, on all available data. The former has not been investigated before in active learning, while the study of latter in the context of deep learning is scarce and recent findings are not conclusive with respect to its benefit. Our idea is orthogonal to acquisition strategies by using more data, much like ensemble methods use more models. By systematically evaluating on a number of popular acquisition strategies and datasets, we find that the use of unlabeled data during model training brings a surprising accuracy improvement in image classification, compared to the differences between acquisition strategies. We thus explore smaller label budgets, even one label per class.

arXiv:1911.08076

IFQ-Net: Integrated Fixed-point Quantization Networks for Embedded Vision

Deploying deep models on embedded devices has been a challenging problem since the great success of deep learning based networks. Fixed-point networks, which represent their data with low bits fixed-point and thus give remarkable savings on memory usage, are generally preferred. Even though current fixed-point networks employ relative low bits (e.g. 8-bits), the memory saving is far from enough for the embedded devices. On the other hand, quantization deep networks, for example XNOR-Net and HWGQNet, quantize the data into 1 or 2 bits resulting in more significant memory savings but still contain lots of floatingpoint data. In this paper, we propose a fixed-point network for embedded vision tasks through converting the floatingpoint data in a quantization network into fixed-point. Furthermore, to overcome the data loss caused by the conversion, we propose to compose floating-point data operations across multiple layers (e.g. convolution, batch normalization and quantization layers) and convert them into fixedpoint. We name the fixed-point network obtained through such integrated conversion as Integrated Fixed-point Quantization Networks (IFQ-Net). We demonstrate that our IFQNet gives 2.16x and 18x more savings on model size and runtime feature map memory respectively with similar accuracy on ImageNet. Furthermore, based on YOLOv2, we design IFQ-Tinier-YOLO face detector which is a fixed-point network with 256x reduction in model size (246k Bytes) than Tiny-YOLO. We illustrate the promising performance of our face detector in terms of detection rate on Face Detection Data Set and Bencmark (FDDB) and qualitative results of detecting small faces of Wider Face dataset.

arXiv:1911.08039공식 코드 · Papers with Code

Reliability Does Matter: An End-to-End Weakly Supervised Semantic Segmentation Approach

Weakly supervised semantic segmentation is a challenging task as it only takes image-level information as supervision for training but produces pixel-level predictions for testing. To address such a challenging task, most recent state-of-the-art approaches propose to adopt two-step solutions, \emph{i.e. } 1) learn to generate pseudo pixel-level masks, and 2) engage FCNs to train the semantic segmentation networks with the pseudo masks. However, the two-step solutions usually employ many bells and whistles in producing high-quality pseudo masks, making this kind of methods complicated and inelegant. In this work, we harness the image-level labels to produce reliable pixel-level annotations and design a fully end-to-end network to learn to predict segmentation maps. Concretely, we firstly leverage an image classification branch to generate class activation maps for the annotated categories, which are further pruned into confident yet tiny object/background regions. Such reliable regions are then directly served as ground-truth labels for the parallel segmentation branch, where a newly designed dense energy loss function is adopted for optimization. Despite its apparent simplicity, our one-step solution achieves competitive mIoU scores (\emph{val}: 62.6, \emph{test}: 62.9) on Pascal VOC compared with those two-step state-of-the-arts. By extending our one-step method to two-step, we get a new state-of-the-art performance on the Pascal VOC (\emph{val}: 66.3, \emph{test}: 66.5).

arXiv:1911.07990공식 코드 · Papers with Code

Crowd Counting via Segmentation Guided Attention Networks and Curriculum Loss

Automatic crowd behaviour analysis is an important task for intelligent transportation systems to enable effective flow control and dynamic route planning for varying road participants. Crowd counting is one of the keys to automatic crowd behaviour analysis. Crowd counting using deep convolutional neural networks (CNN) has achieved encouraging progress in recent years. Researchers have devoted much effort to the design of variant CNN architectures and most of them are based on the pre-trained VGG16 model. Due to the insufficient expressive capacity, the backbone network of VGG16 is usually followed by another cumbersome network specially designed for good counting performance. Although VGG models have been outperformed by Inception models in image classification tasks, the existing crowd counting networks built with Inception modules still only have a small number of layers with basic types of Inception modules. To fill in this gap, in this paper, we firstly benchmark the baseline Inception-v3 model on commonly used crowd counting datasets and achieve surprisingly good performance comparable with or better than most existing crowd counting models. Subsequently, we push the boundary of this disruptive work further by proposing a Segmentation Guided Attention Network (SGANet) with Inception-v3 as the backbone and a novel curriculum loss for crowd counting. We conduct thorough experiments to compare the performance of our SGANet with prior arts and the proposed model can achieve state-of-the-art performance with MAE of 57.6, 6.3 and 87.6 on ShanghaiTechA, ShanghaiTechB and UCF\_QNRF, respectively.

arXiv:1911.07984

Learning Permutation Invariant Representations using Memory Networks

Many real-world tasks such as classification of digital histopathology images and 3D object detection involve learning from a set of instances. In these cases, only a group of instances or a set, collectively, contains meaningful information and therefore only the sets have labels, and not individual data instances. In this work, we present a permutation invariant neural network called Memory-based Exchangeable Model (MEM) for learning set functions. The MEM model consists of memory units that embed an input sequence to high-level features enabling the model to learn inter-dependencies among instances through a self-attention mechanism. We evaluated the learning ability of MEM on various toy datasets, point cloud classification, and classification of lung whole slide images (WSIs) into two subtypes of lung cancer---Lung Adenocarcinoma, and Lung Squamous Cell Carcinoma. We systematically extracted patches from lung WSIs downloaded from The Cancer Genome Atlas~(TCGA) dataset, the largest public repository of WSIs, achieving a competitive accuracy of 84.84\% for classification of two sub-types of lung cancer. The results on other datasets are promising as well, and demonstrate the efficacy of our model.

arXiv:1911.07967

DLBricks: Composable Benchmark Generation to Reduce Deep Learning Benchmarking Effort on CPUs (Extended)

The past few years have seen a surge of applying Deep Learning (DL) models for a wide array of tasks such as image classification, object detection, machine translation, etc. While DL models provide an opportunity to solve otherwise intractable tasks, their adoption relies on them being optimized to meet latency and resource requirements. Benchmarking is a key step in this process but has been hampered in part due to the lack of representative and up-to-date benchmarking suites. This is exacerbated by the fast-evolving pace of DL models. This paper proposes DLBricks, a composable benchmark generation design that reduces the effort of developing, maintaining, and running DL benchmarks on CPUs. DLBricks decomposes DL models into a set of unique runnable networks and constructs the original model's performance using the performance of the generated benchmarks. DLBricks leverages two key observations: DL layers are the performance building blocks of DL models and layers are extensively repeated within and across DL models. Since benchmarks are generated automatically and the benchmarking time is minimized, DLBricks can keep up-to-date with the latest proposed models, relieving the pressure of selecting representative DL models. Moreover, DLBricks allows users to represent proprietary models within benchmark suites. We evaluate DLBricks using $50$ MXNet models spanning $5$ DL tasks on $4$ representative CPU systems. We show that DLBricks provides an accurate performance estimate for the DL models and reduces the benchmarking time across systems (e.g. within $95\%$ accuracy and up to $4.4\times$ benchmarking time speedup on Amazon EC2 c5.xlarge).

arXiv:1911.07940

Neighborhood Watch: Representation Learning with Local-Margin Triplet Loss and Sampling Strategy for K-Nearest-Neighbor Image Classification

Deep representation learning using triplet network for classification suffers from a lack of theoretical foundation and difficulty in tuning both the network and classifiers for performance. To address the problem, local-margin triplet loss along with local positive and negative mining strategy is proposed with theory on how the strategy integrate nearest-neighbor hyper-parameter with triplet learning to increase subsequent classification performance. Results in experiments with 2 public datasets, MNIST and Cifar-10, and 2 small medical image datasets demonstrate that proposed strategy outperforms end-to-end softmax and typical triplet loss in settings without data augmentation while maintaining utility of transferable feature for related tasks. The method serves as a good performance baseline where end-to-end methods encounter difficulties such as small sample data with limited allowable data augmentation.

arXiv:1911.07934

Training Set Effect on Super Resolution for Automated Target Recognition

Single Image Super Resolution (SISR) is the process of mapping a low-resolution image to a high resolution image. This inherently has applications in remote sensing as a way to increase the spatial resolution in satellite imagery. This suggests a possible improvement to automated target recognition in image classification and object detection. We explore the effect that different training sets have on SISR with the network, Super Resolution Generative Adversarial Network (SRGAN). We train 5 SRGANs on different land-use classes (e.g. agriculture, cities, ports) and test them on the same unseen dataset. We attempt to find the qualitative and quantitative differences in SISR, binary classification, and object detection performance. We find that curated training sets that contain objects in the test ontology perform better on both computer vision tasks while having a complex distribution of images allows object detection models to perform better. However, Super Resolution (SR) might not be beneficial to certain problems and will see a diminishing amount of returns for datasets that are closer to being solved.

arXiv:1911.07747

DeepSat V2: Feature Augmented Convolutional Neural Nets for Satellite Image Classification

Satellite image classification is a challenging problem that lies at the crossroads of remote sensing, computer vision, and machine learning. Due to the high variability inherent in satellite data, most of the current object classification approaches are not suitable for handling satellite datasets. The progress of satellite image analytics has also been inhibited by the lack of a single labeled high-resolution dataset with multiple class labels. In a preliminary version of this work, we introduced two new high resolution satellite imagery datasets (SAT-4 and SAT-6) and proposed DeepSat framework for classification based on "handcrafted" features and a deep belief network (DBN). The present paper is an extended version, we present an end-to-end framework leveraging an improved architecture that augments a convolutional neural network (CNN) with handcrafted features (instead of using DBN-based architecture) for classification. Our framework, having access to fused spatial information obtained from handcrafted features as well as CNN feature maps, have achieved accuracies of 99.90% and 99.84% respectively, on SAT-4 and SAT-6, surpassing all the other state-of-the-art results. A statistical analysis based on Distribution Separability Criterion substantiates the robustness of our approach in learning better representations for satellite imagery.