We show that the output of a (residual) convolutional neural network (CNN) with an appropriate prior over the weights and biases is a Gaussian process (GP) in the limit of infinitely many convolutional filters, extending similar results for dense networks. For a CNN, the equivalent kernel can be computed exactly and, unlike "deep kernels", has very few parameters: only the hyperparameters of the original CNN. Further, we show that this kernel has two properties that allow it to be computed efficiently; the cost of evaluating the kernel for a pair of images is similar to a single forward pass through the original CNN with only one filter per layer. The kernel equivalent to a 32-layer ResNet obtains 0.84% classification error on MNIST, a new record for GPs with a comparable number of parameters.
Decentralized machine learning is a promising emerging paradigm in view of global challenges of data ownership and privacy. We consider learning of linear classification and regression models, in the setting where the training data is decentralized over many user devices, and the learning algorithm must run on-device, on an arbitrary communication network, without a central coordinator. We propose COLA, a new decentralized training algorithm with strong theoretical guarantees and superior practical performance. Our framework overcomes many limitations of existing methods, and achieves communication efficiency, scalability, elasticity as well as resilience to changes in data and participating devices.
Learning embedding functions, which map semantically related inputs to nearby locations in a feature space supports a variety of classification and information retrieval tasks. In this work, we propose a novel, generalizable and fast method to define a family of embedding functions that can be used as an ensemble to give improved results. Each embedding function is learned by randomly bagging the training labels into small subsets. We show experimentally that these embedding ensembles create effective embedding functions. The ensemble output defines a metric space that improves state of the art performance for image retrieval on CUB-200-2011, Cars-196, In-Shop Clothes Retrieval and VehicleID.
Convolutional Neural Networks (CNNs) have gained a remarkable success on many image classification tasks in recent years. However, the performance of CNNs highly relies upon their architectures. For most state-of-the-art CNNs, their architectures are often manually-designed with expertise in both CNNs and the investigated problems. Therefore, it is difficult for users, who have no extended expertise in CNNs, to design optimal CNN architectures for their own image classification problems of interest. In this paper, we propose an automatic CNN architecture design method by using genetic algorithms, to effectively address the image classification tasks. The most merit of the proposed algorithm remains in its "automatic" characteristic that users do not need domain knowledge of CNNs when using the proposed algorithm, while they can still obtain a promising CNN architecture for the given images. The proposed algorithm is validated on widely used benchmark image classification datasets, by comparing to the state-of-the-art peer competitors covering eight manually-designed CNNs, seven automatic+manually tuning and five automatic CNN architecture design algorithms. The experimental results indicate the proposed algorithm outperforms the existing automatic CNN architecture design algorithms in terms of classification accuracy, parameter numbers and consumed computational resources. The proposed algorithm also shows the very comparable classification accuracy to the best one from manually-designed and automatic+manually tuning CNNs, while consumes much less of computational resource.
The time series classification literature has expanded rapidly over the last decade, with many new classification approaches published each year. The research focus has mostly been on improving the accuracy and efficiency of classifiers, while their interpretability has been somewhat neglected. Classifier interpretability has become a critical constraint for many application domains and the introduction of the 'right to explanation' GDPR EU legislation in May 2018 is likely to further emphasize the importance of explainable learning algorithms. In this work we analyse the state-of-the-art for time series classification, and propose new algorithms that aim to maintain the classifier accuracy and efficiency, but keep interpretability as a key design constraint. We present new time series classification algorithms that advance the state-of-the-art by implementing the following three key ideas: (1) Multiple resolutions of symbolic approximations: we combine symbolic representations obtained using different parameters; (2) Multiple domain representations: we combine symbolic approximations in time (e.g., SAX) and frequency (e.g., SFA) domains; (3) Efficient navigation of a huge symbolic-words space: we adapt a symbolic sequence classifier named SEQL, to make it work with multiple domain representations (e.g., SAX-SEQL, SFA-SEQL), and use its greedy feature selection strategy to effectively filter the best features for each representation. We show that a multi-resolution multi-domain linear classifier, SAX-SFA-SEQL, achieves a similar accuracy to the state-of-the-art COTE ensemble, and to a recent deep learning method (FCN), but uses a fraction of the time required by either COTE or FCN. We discuss the accuracy, efficiency and interpretability of our proposed algorithms. To further analyse the interpretability aspect of our classifiers, we present a case study on an ecology benchmark.
Computer vision based fine-grained recognition has received great attention in recent years. Existing works focus on discriminative part localization and feature learning. In this paper, to improve the performance of fine-grained recognition, we try to precisely locate as many salient parts of object as possible at first. Then, we figure out the classification probability that can be obtained by using separate parts for object classification. Finally, through extracting efficient features from each part and combining them, then feeding to a classifier for recognition, an improved accuracy over state-of-art algorithms has been obtained on CUB200-2011 bird dataset.
As skin cancer is one of the most frequent cancers globally, accurate, non-invasive dermoscopy-based diagnosis becomes essential and promising. A task of the Part 3 of the ISIC Skin Image Analysis Challenge at MICCAI 2018 is to predict seven disease classes with skin lesion images, including melanoma (MEL), melanocytic nevus (NV), basal cell carcinoma (BCC), actinic keratosis / Bowen's disease (intraepithelial carcinoma) (AKIEC), benign keratosis (solar lentigo / seborrheic keratosis / lichen planus-like keratosis) (BKL), dermatofibroma (DF) and vascular lesion (VASC) as defined by the International Dermatology Society. In this work, we design the WonDerM pipeline, that resamples the preprocessed skin lesion images, builds neural network architecture fine-tuned with segmentation task data (the Part 1), and uses an ensemble method to classify the seven skin diseases. Our model achieved an accuracy of 0.899 and 0.785 in the validation set and test set, respectively.
Characteristics extracted from the training datasets of classification problems have proven to be effective predictors in a number of meta-analyses. Among them, measures of classification complexity can be used to estimate the difficulty in separating the data points into their expected classes. Descriptors of the spatial distribution of the data and estimates of the shape and size of the decision boundary are among the known measures for this characterization. This information can support the formulation of new data-driven pre-processing and pattern recognition techniques, which can in turn be focused on challenges highlighted by such characteristics of the problems. This paper surveys and analyzes measures which can be extracted from the training datasets in order to characterize the complexity of the respective classification problems. Their use in recent literature is also reviewed and discussed, allowing to prospect opportunities for future work in the area. Finally, descriptions are given on an R package named Extended Complexity Library (ECoL) that implements a set of complexity measures and is made publicly available.
Training data plays an essential role in modern applications of machine learning. However, gathering labeled training data is time-consuming. Therefore, labeling is often outsourced to less experienced users, or completely automated. This can introduce errors, which compromise valuable training data, and lead to suboptimal training results. We thus propose a novel approach that uses the power of pretrained classifiers to visually guide users to noisy labels, and let them interactively check error candidates, to iteratively improve the training data set. To systematically investigate training data, we propose a categorization of labeling errors into three different types, based on an analysis of potential pitfalls in label acquisition processes. For each of these types, we present approaches to detect, reason about, and resolve error candidates, as we propose measures and visual guidance techniques to support machine learning users. Our approach has been used to spot errors in well-known machine learning benchmark data sets, and we tested its usability during a user evaluation. While initially developed for images, the techniques presented in this paper are independent of the classification algorithm, and can also be extended to many other types of training data.
Boosting algorithms are frequently used in applied data science and in research. To date, the distinction between boosting with either gradient descent or second-order Newton updates is often not made in both applied and methodological research, and it is thus implicitly assumed that the difference is irrelevant. The goal of this article is to clarify this situation. In particular, we present gradient and Newton boosting, as well as a hybrid variant of the two, in a unified framework. We compare these boosting algorithms with trees as base learners using various datasets and loss functions. Our experiments show that Newton boosting outperforms gradient and hybrid gradient-Newton boosting in terms of predictive accuracy on the majority of datasets. We also present evidence that the reason for this is not faster convergence of Newton boosting. In addition, we introduce a novel tuning parameter for tree-based Newton boosting which is interpretable and important for predictive accuracy.
Data augmentation in deep neural networks is the process of generating artificial data in order to reduce the variance of the classifier with the goal to reduce the number of errors. This idea has been shown to improve deep neural network's generalization capabilities in many computer vision tasks such as image recognition and object localization. Apart from these applications, deep Convolutional Neural Networks (CNNs) have also recently gained popularity in the Time Series Classification (TSC) community. However, unlike in image recognition problems, data augmentation techniques have not yet been investigated thoroughly for the TSC task. This is surprising as the accuracy of deep learning models for TSC could potentially be improved, especially for small datasets that exhibit overfitting, when a data augmentation method is adopted. In this paper, we fill this gap by investigating the application of a recently proposed data augmentation technique based on the Dynamic Time Warping distance, for a deep learning model for TSC. To evaluate the potential of augmenting the training set, we performed extensive experiments using the UCR TSC benchmark. Our preliminary experiments reveal that data augmentation can drastically increase deep CNN's accuracy on some datasets and significantly improve the deep model's accuracy when the method is used in an ensemble approach.
Unsupervised pre-trained word embeddings are used effectively for many tasks in natural language processing to leverage unlabeled textual data. Often these embeddings are either used as initializations or as fixed word representations for task-specific classification models. In this work, we extend our classification model's task loss with an unsupervised auxiliary loss on the word-embedding level of the model. This is to ensure that the learned word representations contain both task-specific features, learned from the supervised loss component, and more general features learned from the unsupervised loss component. We evaluate our approach on the task of temporal relation extraction, in particular, narrative containment relation extraction from clinical records, and show that continued training of the embeddings on the unsupervised objective together with the task objective gives better task-specific embeddings, and results in an improvement over the state of the art on the THYME dataset, using only a general-domain part-of-speech tagger as linguistic resource.
The prediction accuracy has been the long-lasting and sole standard for comparing the performance of different image classification models, including the ImageNet competition. However, recent studies have highlighted the lack of robustness in well-trained deep neural networks to adversarial examples. Visually imperceptible perturbations to natural images can easily be crafted and mislead the image classifiers towards misclassification. To demystify the trade-offs between robustness and accuracy, in this paper we thoroughly benchmark 18 ImageNet models using multiple robustness metrics, including the distortion, success rate and transferability of adversarial examples between 306 pairs of models. Our extensive experimental results reveal several new insights: (1) linear scaling law - the empirical $\ell_2$ and $\ell_\infty$ distortion metrics scale linearly with the logarithm of classification error; (2) model architecture is a more critical factor to robustness than model size, and the disclosed accuracy-robustness Pareto frontier can be used as an evaluation criterion for ImageNet model designers; (3) for a similar network architecture, increasing network depth slightly improves robustness in $\ell_\infty$ distortion; (4) there exist models (in VGG family) that exhibit high adversarial transferability, while most adversarial examples crafted from one model can only be transferred within the same family. Experiment code is publicly available at \url{https://github.com/huanzhang12/Adversarial_Survey}.
In recent years there has been a sharp rise in networking applications, in which significant events need to be classified but only a few training instances are available. These are known as cases of one-shot learning. Examples include analyzing network traffic under zero-day attacks, and computer vision tasks by sensor networks deployed in the field. To handle this challenging task, organizations often use human analysts to classify events under high uncertainty. Existing algorithms use a threshold-based mechanism to decide whether to classify an object automatically or send it to an analyst for deeper inspection. However, this approach leads to a significant waste of resources since it does not take the practical temporal constraints of system resources into account. Our contribution is threefold. First, we develop a novel Deep Reinforcement One-shot Learning (DeROL) framework to address this challenge. The basic idea of the DeROL algorithm is to train a deep-Q network to obtain a policy which is oblivious to the unseen classes in the testing data. Then, in real-time, DeROL maps the current state of the one-shot learning process to operational actions based on the trained deep-Q network, to maximize the objective function. Second, we develop the first open-source software for practical artificially intelligent one-shot classification systems with limited resources for the benefit of researchers in related fields. Third, we present an extensive experimental study using the OMNIGLOT dataset for computer vision tasks and the UNSW-NB15 dataset for intrusion detection tasks that demonstrates the versatility and efficiency of the DeROL framework.
In generalized zero shot learning (GZSL), the set of classes are split into seen and unseen classes, where training relies on the semantic features of the seen and unseen classes and the visual representations of only the seen classes, while testing uses the visual representations of the seen and unseen classes. Current methods address GZSL by learning a transformation from the visual to the semantic space, exploring the assumption that the distribution of classes in the semantic and visual spaces is relatively similar. Such methods tend to transform unseen testing visual representations into one of the seen classes' semantic features instead of the semantic features of the correct unseen class, resulting in low accuracy GZSL classification. Recently, generative adversarial networks (GAN) have been explored to synthesize visual representations of the unseen classes from their semantic features - the synthesized representations of the seen and unseen classes are then used to train the GZSL classifier. This approach has been shown to boost GZSL classification accuracy, however, there is no guarantee that synthetic visual representations can generate back their semantic feature in a multi-modal cycle-consistent manner. This constraint can result in synthetic visual representations that do not represent well their semantic features. In this paper, we propose the use of such constraint based on a new regularization for the GAN training that forces the generated visual features to reconstruct their original semantic features. Once our model is trained with this multi-modal cycle-consistent semantic compatibility, we can then synthesize more representative visual representations for the seen and, more importantly, for the unseen classes. Our proposed approach shows the best GZSL classification results in the field in several publicly available datasets.
While implicit generative models such as GANs have shown impressive results in high quality image reconstruction and manipulation using a combination of various losses, we consider a simpler approach leading to surprisingly strong results. We show that texture loss alone allows the generation of perceptually high quality images. We provide a better understanding of texture constraining mechanism and develop a novel semantically guided texture constraining method for further improvement. Using a recently developed perceptual metric employing "deep features" and termed LPIPS, the method obtains state-of-the-art results. Moreover, we show that a texture representation of those deep features better capture the perceptual quality of an image than the original deep features. Using texture information, off-the-shelf deep classification networks (without training) perform as well as the best performing (tuned and calibrated) LPIPS metrics. The code is publicly available.
Modern CNN-based object detectors rely on bounding box regression and non-maximum suppression to localize objects. While the probabilities for class labels naturally reflect classification confidence, localization confidence is absent. This makes properly localized bounding boxes degenerate during iterative regression or even suppressed during NMS. In the paper we propose IoU-Net learning to predict the IoU between each detected bounding box and the matched ground-truth. The network acquires this confidence of localization, which improves the NMS procedure by preserving accurately localized bounding boxes. Furthermore, an optimization-based bounding box refinement method is proposed, where the predicted IoU is formulated as the objective. Extensive experiments on the MS-COCO dataset show the effectiveness of IoU-Net, as well as its compatibility with and adaptivity to several state-of-the-art object detectors.
We study the problem of unsupervised domain adaptive re-identification (re-ID) which is an active topic in computer vision but lacks a theoretical foundation. We first extend existing unsupervised domain adaptive classification theories to re-ID tasks. Concretely, we introduce some assumptions on the extracted feature space and then derive several loss functions guided by these assumptions. To optimize them, a novel self-training scheme for unsupervised domain adaptive re-ID tasks is proposed. It iteratively makes guesses for unlabeled target data based on an encoder and trains the encoder based on the guessed labels. Extensive experiments on unsupervised domain adaptive person re-ID and vehicle re-ID tasks with comparisons to the state-of-the-arts confirm the effectiveness of the proposed theories and self-training framework. Our code is available at \url{https://github.com/LcDog/DomainAdaptiveReID}.
Most activity localization methods in the literature suffer from the burden of frame-wise annotation requirement. Learning from weak labels may be a potential solution towards reducing such manual labeling effort. Recent years have witnessed a substantial influx of tagged videos on the Internet, which can serve as a rich source of weakly-supervised training data. Specifically, the correlations between videos with similar tags can be utilized to temporally localize the activities. Towards this goal, we present W-TALC, a Weakly-supervised Temporal Activity Localization and Classification framework using only video-level labels. The proposed network can be divided into two sub-networks, namely the Two-Stream based feature extractor network and a weakly-supervised module, which we learn by optimizing two complimentary loss functions. Qualitative and quantitative results on two challenging datasets - Thumos14 and ActivityNet1.2, demonstrate that the proposed method is able to detect activities at a fine granularity and achieve better performance than current state-of-the-art methods.
Hierarchical Bayesian methods can unify many related tasks (e.g. k-shot classification, conditional and unconditional generation) as inference within a single generative model. However, when this generative model is expressed as a powerful neural network such as a PixelCNN, we show that existing learning techniques typically fail to effectively use latent variables. To address this, we develop a modification of the Variational Autoencoder in which encoded observations are decoded to new elements from the same class. This technique, which we call a Variational Homoencoder (VHE), produces a hierarchical latent variable model which better utilises latent variables. We use the VHE framework to learn a hierarchical PixelCNN on the Omniglot dataset, which outperforms all existing models on test set likelihood and achieves strong performance on one-shot generation and classification tasks. We additionally validate the VHE on natural images from the YouTube Faces database. Finally, we develop extensions of the model that apply to richer dataset structures such as factorial and hierarchical categories.