Convolutional neural networks (CNNs) are commonly trained using a fixed spatial image size predetermined for a given model. Although trained on images of aspecific size, it is well established that CNNs can be used to evaluate a wide range of image sizes at test time, by adjusting the size of intermediate feature maps. In this work, we describe and evaluate a novel mixed-size training regime that mixes several image sizes at training time. We demonstrate that models trained using our method are more resilient to image size changes and generalize well even on small images. This allows faster inference by using smaller images attest time. For instance, we receive a 76.43% top-1 accuracy using ResNet50 with an image size of 160, which matches the accuracy of the baseline model with 2x fewer computations. Furthermore, for a given image size used at test time, we show this method can be exploited either to accelerate training or the final test accuracy. For example, we are able to reach a 79.27% accuracy with a model evaluated at a 288 spatial size for a relative improvement of 14% over the baseline.

In E-commerce, it is a common practice to organize the product catalog using product taxonomy. This enables the buyer to easily locate the item they are looking for and also to explore various items available under a category. Product taxonomy is a tree structure with 3 or more levels of depth and several leaf nodes. Product categorization is a large scale classification task that assigns a category path to a particular product. Research in this area is restricted by the unavailability of good real-world datasets and the variations in taxonomy due to the absence of a standard across the different e-commerce stores. In this paper, we introduce a high-quality product taxonomy dataset focusing on clothing products which contain 186,150 images under clothing category with 3 levels and 52 leaf nodes in the taxonomy. We explain the methodology used to collect and label this dataset. Further, we establish the benchmark by comparing image classification and Attention based Sequence models for predicting the category path. Our benchmark model reaches a micro f-score of 0.92 on the test set. The dataset, code and pre-trained models are publicly available at \url{https://github.com/vumaasha/atlas}. We invite the community to improve upon these baselines.

Figures are an important channel for scientific communication, used to express complex ideas, models and data in ways that words cannot. However, this visual information is mostly ignored in analyses of the scientific literature. In this paper, we demonstrate the utility of using scientific figures as markers of knowledge domains in science, which can be used for classification, recommender systems, and studies of scientific information exchange. We encode sets of images into a visual signature, then use distances between these signatures to understand how patterns of visual communication compare with patterns of jargon and citation structures. We find that figures can be as effective for differentiating communities of practice as text or citation patterns. We then consider where these metrics disagree to understand how different disciplines use visualization to express ideas. Finally, we further consider how specific figure types propagate through the literature, suggesting a new mechanism for understanding the flow of ideas apart from conventional channels of text and citations. Our ultimate aim is to better leverage these information-dense objects to improve scientific communication across disciplinary boundaries.

Tensor decomposition methods are one of the primary approaches for model compression and fast inference of convolutional neural networks (CNNs). However, despite their potential diversity, only a few typical decompositions such as CP decomposition have been applied in practice; more importantly, no extensive comparisons have been performed between available methods. This raises the simple question of how many decompositions are possible, and which of these is the best. In this paper, we first characterize a decomposition class specific to CNNs by adopting graphical notation, which is considerably flexible. When combining with the nonlinear activations, the class includes renowned CNN modules such as depthwise separable convolution and bottleneck layer. In the experiments, we compare the tradeoff between prediction accuracy and time/space complexities by enumerating all the possible decompositions. Also, we demonstrate, using a neural architecture search, that we can find nonlinear decompositions that outperform existing decompositions.

Neural text generation is a key tool in natural language applications, but it is well known there are major problems at its core. In particular, standard likelihood training and decoding leads to dull and repetitive responses. While some post-hoc fixes have been proposed, in particular top-k and nucleus sampling, they do not address the fact that the token-level probabilities predicted by the model itself are poor. In this paper we show that the likelihood objective itself is at fault, resulting in a model that assigns too much probability to sequences that contain repeats and frequent words unlike the human training distribution. We propose a new objective, unlikelihood training, which forces unlikely generations to be assigned lower probability by the model. We show that both token and sequence level unlikelihood training give less repetitive, less dull text while maintaining perplexity, giving far superior generations using standard greedy or beam search. Our approach provides a strong alternative to traditional training.

Batch Normalization (BN) (Ioffe and Szegedy 2015) normalizes the features of an input image via statistics of a batch of images and this batch information is considered as batch noise that will be brought to the features of an instance by BN. We offer a point of view that self-attention mechanism can help regulate the batch noise by enhancing instance-specific information. Based on this view, we propose combining BN with a self-attention mechanism to adjust the batch noise and give an attention-based version of BN called Instance Enhancement Batch Normalization (IEBN) which recalibrates channel information by a simple linear transformation. IEBN outperforms BN with a light parameter increment in various visual tasks universally for different network structures and benchmark data sets. Besides, even if under the attack of synthetic noise, IEBN can still stabilize network training with good generalization. The code of IEBN is available at https://github.com/gbup-group/IEBN

We demonstrate the existence of universal adversarial perturbations, which can fool a family of audio processing architectures, for both targeted and untargeted attacks. To the best of our knowledge, this is the first study on generating universal adversarial perturbations for audio processing systems. We propose two methods for finding such perturbations. The first method is based on an iterative, greedy approach that is well-known in computer vision: it aggregates small perturbations to the input so as to push it to the decision boundary. The second method, which is the main technical contribution of this work, is a novel penalty formulation, which finds targeted and untargeted universal adversarial perturbations. Differently from the greedy approach, the penalty method minimizes an appropriate objective function on a batch of samples. Therefore, it produces more successful attacks when the number of training samples is limited. Moreover, we provide a proof that the proposed penalty method theoretically converges to a solution that corresponds to universal adversarial perturbations. We report comprehensive experiments, showing attack success rates higher than 91.1% and 74.7% for targeted and untargeted attacks, respectively.

Weakly supervised object localization (WSOL) models aim to locate objects of interest in an image after being trained only on data with coarse image level labels. Deep learning models for WSOL rely typically on convolutional attention maps with no constraints on the regions of interest which allows them to select any region, making them vulnerable to false positive regions. This issue occurs in many application domains, e.g., medical image analysis, where interpretability is central to the prediction. In order to improve the localization reliability, we propose a deep learning framework for WSOL with pixel level localization. It is composed of two sequential sub-networks: a localizer that localizes regions of interest; followed by a classifier that classifies them. Within its end-to-end training, we incorporate the prior knowledge stating that in an agnostic-class setup an image is more likely to contain relevant --object of interest-- and irrelevant regions --noise--. Based on the conditional entropy (CE) measured at the classifier, the localizer is driven to spot relevant regions (low CE), and irrelevant regions (high CE). Our framework is able to recover large discriminative regions using our recursive erasing algorithm that we incorporate within the backpropagation during training. Moreover, the framework handles intrinsically multi-instances. Experimental results on public datasets with medical images (GlaS colon cancer) and natural images (Caltech-UCSD Birds-200-2011, Oxford flower 102) show that, compared to state of the art WSOL methods, our framework can provide significant improvements in terms of image-level classification, pixel-level localization, and robustness to overfitting when dealing with few training samples. A public reproducible PyTorch implementation is provided in: https://github.com/sbelharbi/wsol-min-max-entropy-interpretability .

Building a Massive Corpus for Named Entity Recognition using Free Open Data Sources Daniel Specht Menezes Departamento de Inform atica PUC-Rio Rio de Janeiro, Brazil danielssmenezes@gmail.com Abstract --With the recent progress in machine learning, boosted by techniques such as deep learning, many tasks can be successfully solved once a large enough dataset is available for training. Nonetheless, human-annotated datasets are often expensive to produce, especially when labels are fine-grained, as is the case of Named Entity Recognition (NER), a task that operates with labels on a word-level. In this paper, we propose a method to automatically generate labeled datasets for NER from public data sources by exploiting links and structured data from DBpedia and Wikipedia. Due to the massive size of these data sources, the resulting dataset - SESAME 1 - is composed of millions of labeled sentences. We detail the method to generate the dataset, report relevant statistics, and design a baseline using a neural network, showing that our dataset helps building better NER predictors.

Using class labels to represent class similarity is a typical approach to training deep hashing systems for retrieval; samples from the same or different classes take binary 1 or 0 similarity values. This similarity does not model the full rich knowledge of semantic relations that may be present between data points. In this work we build upon the idea of using semantic hierarchies to form distance metrics between all available sample labels; for example cat to dog has a smaller distance than cat to guitar. We combine this type of semantic distance into a loss function to promote similar distances between the deep neural network embeddings. We also introduce an empirical Kullback-Leibler divergence loss term to promote binarization and uniformity of the embeddings. We test the resulting SHREWD method and demonstrate improvements in hierarchical retrieval scores using compact, binary hash codes instead of real valued ones, and show that in a weakly supervised hashing setting we are able to learn competitively without explicitly relying on class labels, but instead on similarities between labels.