In this paper, we introduce a novel methodology for characterising the performance of deep learning networks (ResNets and DenseNet) with respect to training convergence and generalisation as a function of mini-batch size and learning rate for image classification. This methodology is based on novel measurements derived from the eigenvalues of the approximate Fisher information matrix, which can be efficiently computed even for high capacity deep models. Our proposed measurements can help practitioners to monitor and control the training process (by actively tuning the mini-batch size and learning rate) to allow for good training convergence and generalisation. Furthermore, the proposed measurements also allow us to show that it is possible to optimise the training process with a new dynamic sampling training approach that continuously and automatically change the mini-batch size and learning rate during the training process. Finally, we show that the proposed dynamic sampling training approach has a faster training time and a competitive classification accuracy compared to the current state of the art.
Hazy images are common in real scenarios and many dehazing methods have been developed to automatically remove the haze from images. Typically, the goal of image dehazing is to produce clearer images from which human vision can better identify the object and structural details present in the images. When the ground-truth haze-free image is available for a hazy image, quantitative evaluation of image dehazing is usually based on objective metrics, such as Peak Signal-to-Noise Ratio (PSNR) and Structural Similarity (SSIM). However, in many applications, large-scale images are collected not for visual examination by human. Instead, they are used for many high-level vision tasks, such as automatic classification, recognition and categorization. One fundamental problem here is whether various dehazing methods can produce clearer images that can help improve the performance of the high-level tasks. In this paper, we empirically study this problem in the important task of image classification by using both synthetic and real hazy image datasets. From the experimental results, we find that the existing image-dehazing methods cannot improve much the image-classification performance and sometimes even reduce the image-classification performance.
Compressive Sensing (CS) theory asserts that sparse signal reconstruction is possible from a small number of linear measurements. Although CS enables low-cost linear sampling, it requires non-linear and costly reconstruction. Recent literature works show that compressive image classification is possible in CS domain without reconstruction of the signal. In this work, we introduce a DCT base method that extracts binary discriminative features directly from CS measurements. These CS measurements can be obtained by using (i) a random or a pseudo-random measurement matrix, or (ii) a measurement matrix whose elements are learned from the training data to optimize the given classification task. We further introduce feature fusion by concatenating Bag of Words (BoW) representation of our binary features with one of the two state-of-the-art CNN-based feature vectors. We show that our fused feature outperforms the state-of-the-art in both cases.
Designing the architecture for an artificial neural network is a cumbersome task because of the numerous parameters to configure, including activation functions, layer types, and hyper-parameters. With the large number of parameters for most networks nowadays, it is intractable to find a good configuration for a given task by hand. In this paper an Efficient Global Optimization (EGO) algorithm is adapted to automatically optimize and configure convolutional neural network architectures. A configurable neural network architecture based solely on convolutional layers is proposed for the optimization. Without using any knowledge on the target problem and not using any data augmentation techniques, it is shown that on several image classification tasks this approach is able to find competitive network architectures in terms of prediction accuracy, compared to the best hand-crafted ones in literature. In addition, a very small training budget (200 evaluations and 10 epochs in training) is spent on each optimized architectures in contrast to the usual long training time of hand-crafted networks. Moreover, instead of the standard sequential evaluation in EGO, several candidate architectures are proposed and evaluated in parallel, which saves the execution overheads significantly and leads to an efficient automation for deep neural network design.
Just like many other topics in computer vision, image classification has achieved significant progress recently by using deep-learning neural networks, especially the Convolutional Neural Networks (CNN). Most of the existing works are focused on classifying very clear natural images, evidenced by the widely used image databases such as Caltech-256, PASCAL VOCs and ImageNet. However, in many real applications, the acquired images may contain certain degradations that lead to various kinds of blurring, noise, and distortions. One important and interesting problem is the effect of such degradations to the performance of CNN-based image classification. More specifically, we wonder whether image-classification performance drops with each kind of degradation, whether this drop can be avoided by including degraded images into training, and whether existing computer vision algorithms that attempt to remove such degradations can help improve the image-classification performance. In this paper, we empirically study this problem for four kinds of degraded images -- hazy images, underwater images, motion-blurred images and fish-eye images. For this study, we synthesize a large number of such degraded images by applying respective physical models to the clear natural images and collect a new hazy image dataset from the Internet. We expect this work can draw more interests from the community to study the classification of degraded images.
Understanding the uncertainty of a neural network's (NN) predictions is essential for many purposes. The Bayesian framework provides a principled approach to this, however applying it to NNs is challenging due to large numbers of parameters and data. Ensembling NNs provides an easily implementable, scalable method for uncertainty quantification, however, it has been criticised for not being Bayesian. This work proposes one modification to the usual process that we argue does result in approximate Bayesian inference; regularising parameters about values drawn from a distribution which can be set equal to the prior. A theoretical analysis of the procedure in a simplified setting suggests the recovered posterior is centred correctly but tends to have an underestimated marginal variance, and overestimated correlation. However, two conditions can lead to exact recovery. We argue that these conditions are partially present in NNs. Empirical evaluations demonstrate it has an advantage over standard ensembling, and is competitive with variational methods.
In contrast to the standard classification paradigm where the true class is given to each training pattern, complementary-label learning only uses training patterns each equipped with a complementary label, which only specifies one of the classes that the pattern does not belong to. The goal of this paper is to derive a novel framework of complementary-label learning with an unbiased estimator of the classification risk, for arbitrary losses and models---all existing methods have failed to achieve this goal. Not only is this beneficial for the learning stage, it also makes model/hyper-parameter selection (through cross-validation) possible without the need of any ordinarily labeled validation data, while using any linear/non-linear models or convex/non-convex loss functions. We further improve the risk estimator by a non-negative correction and gradient ascent trick, and demonstrate its superiority through experiments.
In this brief technical report we introduce the CINIC-10 dataset as a plug-in extended alternative for CIFAR-10. It was compiled by combining CIFAR-10 with images selected and downsampled from the ImageNet database. We present the approach to compiling the dataset, illustrate the example images for different classes, give pixel distributions for each part of the repository, and give some standard benchmarks for well known models. Details for download, usage, and compilation can be found in the associated github repository.
Convolutional Neural Networks (CNNs) are a class of artificial neural networks whose computational blocks use convolution, together with other linear and non-linear operations, to perform classification or regression. This paper explores the spectral response of CNNs and its potential use in diagnosing problems with their training. We measure the gain of CNNs trained for image classification on ImageNet and observe that the best models are also the most sensitive to perturbations of their input. Further, we perform experiments on MNIST and CIFAR-10 to find that the gain rises as the network learns and then saturates as the network converges. Moreover, we find that strong gain fluctuations can point to overfitting and learning problems caused by a poor choice of learning rate. We argue that the gain of CNNs can act as a diagnostic tool and potential replacement for the validation loss when hold-out validation data are not available.
We explore the generalization of scattering transforms from traditional (e.g., image or audio) signals to graph data, analogous to the generalization of ConvNets in geometric deep learning, and the utility of extracted graph features in graph data analysis. In particular, we focus on the capacity of these features to retain informative variability and relations in the data (e.g., between individual graphs, or in aggregate), while relating our construction to previous theoretical results that establish the stability of similar transforms to families of graph deformations. We demonstrate the application the our geometric scattering features in graph classification of social network data, and in data exploration of biochemistry data.
We propose deep convolutional Gaussian processes, a deep Gaussian process architecture with convolutional structure. The model is a principled Bayesian framework for detecting hierarchical combinations of local features for image classification. We demonstrate greatly improved image classification performance compared to current Gaussian process approaches on the MNIST and CIFAR-10 datasets. In particular, we improve CIFAR-10 accuracy by over 10 percentage points.
Pruning large neural networks while maintaining their performance is often desirable due to the reduced space and time complexity. In existing methods, pruning is done within an iterative optimization procedure with either heuristically designed pruning schedules or additional hyperparameters, undermining their utility. In this work, we present a new approach that prunes a given network once at initialization prior to training. To achieve this, we introduce a saliency criterion based on connection sensitivity that identifies structurally important connections in the network for the given task. This eliminates the need for both pretraining and the complex pruning schedule while making it robust to architecture variations. After pruning, the sparse network is trained in the standard way. Our method obtains extremely sparse networks with virtually the same accuracy as the reference network on the MNIST, CIFAR-10, and Tiny-ImageNet classification tasks and is broadly applicable to various architectures including convolutional, residual and recurrent networks. Unlike existing methods, our approach enables us to demonstrate that the retained connections are indeed relevant to the given task.
The low displacement rank (LDR) framework for structured matrices represents a matrix through two displacement operators and a low-rank residual. Existing use of LDR matrices in deep learning has applied fixed displacement operators encoding forms of shift invariance akin to convolutions. We introduce a class of LDR matrices with more general displacement operators, and explicitly learn over both the operators and the low-rank component. This class generalizes several previous constructions while preserving compression and efficient computation. We prove bounds on the VC dimension of multi-layer neural networks with structured weight matrices and show empirically that our compact parameterization can reduce the sample complexity of learning. When replacing weight layers in fully-connected, convolutional, and recurrent neural networks for image classification and language modeling tasks, our new classes exceed the accuracy of existing compression approaches, and on some tasks also outperform general unstructured layers while using more than 20x fewer parameters.
Bag-of-Visual-Words (BoVW) approach has been widely used in the recent years for image classification purposes. However, the limitations regarding optimal feature selection, clustering technique, the lack of spatial organization of the data and the weighting of visual words are crucial. These factors affect the stability of the model and reduce performance. We propose to develop an algorithm based on BoVW for facial expression analysis which goes beyond those limitations. Thus the visual codebook is built by using k-Means++ method to avoid poor clustering. To exploit reliable low level features, we search for the best feature detector that avoids locating a large number of keypoints which do not contribute to the classification process. Then, we propose to compute the relative conjunction matrix in order to preserve the spatial order of the data by coding the relationships among visual words. In addition, a weighting scheme that reflects how important a visual word is with respect to a given image is introduced. We speed up the learning process by using histogram intersection kernel by Support Vector Machine to learn a discriminative classifier. The efficiency of the proposed algorithm is compared with standard bag of visual words method and with bag of visual words method with spatial pyramid. Extensive experiments on the CK+, the MMI and the JAFFE databases show good average recognition rates. Likewise, the ability to recognize spontaneous and non-basic expressive states is investigated using the DynEmo database.
This paper presents an automatic network adaptation method that finds a ConvNet structure well-suited to a given target task, e.g., image classification, for efficiency as well as accuracy in transfer learning. We call the concept target-aware transfer learning. Given only small-scale labeled data, and starting from an ImageNet pre-trained network, we exploit a scheme of removing its potential redundancy for the target task through iterative operations of filter-wise pruning and network optimization. The basic motivation is that compact networks are on one hand more efficient and should also be more tolerant, being less complex, against the risk of overfitting which would hinder the generalization of learned representations in the context of transfer learning. Further, unlike existing methods involving network simplification, we also let the scheme identify redundant portions across the entire network, which automatically results in a network structure adapted to the task at hand. We achieve this with a few novel ideas: (i) cumulative sum of activation statistics for each layer, and (ii) a priority evaluation of pruning across multiple layers. Experimental results by the method on five datasets (Flower102, CUB200-2011, Dog120, MIT67, and Stanford40) show favorable accuracies over the related state-of-the-art techniques while enhancing the computational and storage efficiency of the transferred model.
Recognizing the type of an ancient coin requires theoretical expertise and years of experience in the field of numismatics. Our goal in this work is automatizing this time consuming and demanding task by a visual classification framework. Specifically, we propose to model ancient coin image classification using Graph Transduction Games (GTG). GTG casts the classification problem as a non-cooperative game where the players (the coin images) decide their strategies (class labels) according to the choices made by the others, which results with a global consensus at the final labeling. Experiments are conducted on the only publicly available dataset which is composed of 180 images of 60 types of Roman coins. We demonstrate that our approach outperforms the literature work on the same dataset with the classification accuracy of 73.6% and 87.3% when there are one and two images per class in the training set, respectively.
Many machine learning tasks such as multiple instance learning, 3D shape recognition, and few-shot image classification are defined on sets of instances. Since solutions to such problems do not depend on the order of elements of the set, models used to address them should be permutation invariant. We present an attention-based neural network module, the Set Transformer, specifically designed to model interactions among elements in the input set. The model consists of an encoder and a decoder, both of which rely on attention mechanisms. In an effort to reduce computational complexity, we introduce an attention scheme inspired by inducing point methods from sparse Gaussian process literature. It reduces the computation time of self-attention from quadratic to linear in the number of elements in the set. We show that our model is theoretically attractive and we evaluate it on a range of tasks, demonstrating the state-of-the-art performance compared to recent methods for set-structured data.
In this paper, we target the problem of fracture classification from clinical X-Ray images towards an automated Computer Aided Diagnosis (CAD) system. Although primarily dealing with an image classification problem, we argue that localizing the fracture in the image is crucial to make good class predictions. Therefore, we propose and thoroughly analyze several schemes for simultaneous fracture localization and classification. We show that using an auxiliary localization task, in general, improves the classification performance. Moreover, it is possible to avoid the need for additional localization annotations thanks to recent advancements in weakly-supervised deep learning approaches. Among such approaches, we investigate and adapt Spatial Transformers (ST), Self-Transfer Learning (STL), and localization from global pooling layers. We provide a detailed quantitative and qualitative validation on a dataset of 1347 femur fractures images and report high accuracy with regard to inter-expert correlation values reported in the literature. Our investigations show that i) lesion localization improves the classification outcome, ii) weakly-supervised methods improve baseline classification without any additional cost, iii) STL guides feature activations and boost performance. We plan to make both the dataset and code available.
We propose an efficient way to output better calibrated uncertainty scores from neural networks. The Distilled Dropout Network (DDN) makes standard (non-Bayesian) neural networks more introspective by adding a new training loss which prevents them from being overconfident. Our method is more efficient than Bayesian neural networks or model ensembles which, despite providing more reliable uncertainty scores, are more cumbersome to train and slower to test. We evaluate DDN on the the task of image classification on the CIFAR-10 dataset and show that our calibration results are competitive even when compared to 100 Monte Carlo samples from a dropout network while they also increase the classification accuracy. We also propose better calibration within the state of the art Faster R-CNN object detection framework and show, using the COCO dataset, that DDN helps train better calibrated object detectors.
Recent work has studied the reasons for the remarkable performance of deep neural networks in image classification. We examine batch normalization on the one hand and the dynamical systems view of residual networks on the other hand. Our goal is in understanding the notions of stability and smoothness of the inter-layer propagation of ResNets so as to explain when they contribute to significantly enhanced performance. We postulate that such stability is of importance for the trained ResNet to transfer.