Amazon (Germany)

Much of the recent progress made in image classification research can be credited to training procedure refinements, such as changes in data augmentations and optimization methods. In the literature, however, most refinements are either briefly mentioned as implementation details or only visible in source code. In this paper, we will examine a collection of such refinements and empirically evaluate their impact on the final model accuracy through ablation study. We will show that, by combining these refinements together, we are able to improve various CNN models significantly. For example, we raise ResNet-50's top-1 validation accuracy from 75.3% to 79.29% on ImageNet. We will also demonstrate that improvement on image classification accuracy leads to better transfer learning performance in other application domains such as object detection and semantic segmentation.

The ability to learn richer network representations generally boosts the performance of deep learning models. To improve representation-learning in convolutional neural networks, we present a multi-branch architecture, which applies channel-wise attention across different network branches to leverage the complementary strengths of both feature-map attention and multi-path representation. Our proposed Split-Attention module provides a simple and modular computation block that can serve as a drop-in replacement for the popular residual block, while producing more diverse representations via cross-feature interactions. Adding a Split-Attention module into the architecture design space of RegNet-Y and FBNetV2 directly improves the performance of the resulting network. Replacing residual blocks with our Split-Attention module, we further design a new variant of the ResNet model, named ResNeSt, which outperforms EfficientNet in terms of the accuracy/latency trade-off.

Building reliable distributed systems at a worldwide scale demands trade-offs between consistency and availability.

Deep embeddings answer one simple question: How similar are two images? Learning these embeddings is the bedrock of verification, zero-shot learning, and visual search. The most prominent approaches optimize a deep convolutional network with a suitable loss function, such as contrastive loss or triplet loss. While a rich line of work focuses solely on the loss functions, we show in this paper that selecting training examples plays an equally important role. We propose distance weighted sampling, which selects more informative and stable examples than traditional approaches. In addition, we show that a simple margin based loss is sufficient to outperform all other loss functions. We evaluate our approach on the Stanford Online Products, CAR196, and the CUB200-2011 datasets for image retrieval and clustering, and on the LFW dataset for face verification. Our method achieves state-of-the-art performance on all of them.

We present our experience with QUIC, an encrypted, multiplexed, and low-latency transport protocol designed from the ground up to improve transport performance for HTTPS traffic and to enable rapid deployment and continued evolution of transport mechanisms. QUIC has been globally deployed at Google on thousands of servers and is used to serve traffic to a range of clients including a widely-used web browser (Chrome) and a popular mobile video streaming app (YouTube). We estimate that 7% of Internet traffic is now QUIC. We describe our motivations for developing a new transport, the principles that guided our design, the Internet-scale process that we used to perform iterative experiments on QUIC, performance improvements seen by our various services, and our experience deploying QUIC globally. We also share lessons about transport design and the Internet ecosystem that we learned from our deployment.

Region anchors are the cornerstone of modern object detection techniques. State-of-the-art detectors mostly rely on a dense anchoring scheme, where anchors are sampled uniformly over the spatial domain with a predefined set of scales and aspect ratios. In this paper, we revisit this foundational stage. Our study shows that it can be done much more effectively and efficiently. Specifically, we present an alternative scheme, named Guided Anchoring, which leverages semantic features to guide the anchoring. The proposed method jointly predicts the locations where the center of objects of interest are likely to exist as well as the scales and aspect ratios at different locations. On top of predicted anchor shapes, we mitigate the feature inconsistency with a feature adaption module. We also study the use of high-quality proposals to improve detection performance. The anchoring scheme can be seamlessly integrated into proposal methods and detectors. With Guided Anchoring, we achieve 9.1% higher recall on MS COCO with 90% fewer anchors than the RPN baseline. We also adopt Guided Anchoring in Fast R-CNN, Faster R-CNN and RetinaNet, respectively improving the detection mAP by 2.2%, 2.7% and 1.2%. Code is available at https://github.com/open-mmlab/mmdetection.

Polygonal meshes provide an efficient representation for 3D shapes. They explicitly captureboth shape surface and topology, and leverage non-uniformity to represent large flat regions as well as sharp, intricate features. This non-uniformity and irregularity, however, inhibits mesh analysis efforts using neural networks that combine convolution and pooling operations. In this paper, we utilize the unique properties of the mesh for a direct analysis of 3D shapes using MeshCNN , a convolutional neural network designed specifically for triangular meshes. Analogous to classic CNNs, MeshCNN combines specialized convolution and pooling layers that operate on the mesh edges, by leveraging their intrinsic geodesic connections. Convolutions are applied on edges and the four edges of their incident triangles, and pooling is applied via an edge collapse operation that retains surface topology, thereby, generating new mesh connectivity for the subsequent convolutions. MeshCNN learns which edges to collapse, thus forming a task-driven process where the network exposes and expands the important features while discarding the redundant ones. We demonstrate the effectiveness of MeshCNN on various learning tasks applied to 3D meshes.

Transferring knowledge from a teacher neural network pretrained on the same or a similar task to a student neural network can significantly improve the performance of the student neural network. Existing knowledge transfer approaches match the activations or the corresponding hand-crafted features of the teacher and the student networks. We propose an information-theoretic framework for knowledge transfer which formulates knowledge transfer as maximizing the mutual information between the teacher and the student networks. We compare our method with existing knowledge transfer methods on both knowledge distillation and transfer learning tasks and show that our method consistently outperforms existing methods. We further demonstrate the strength of our method on knowledge transfer across heterogeneous network architectures by transferring knowledge from a convolutional neural network (CNN) to a multi-layer perceptron (MLP) on CIFAR-10. The resulting MLP significantly outperforms the-state-of-the-art methods and it achieves similar performance to the CNN with a single convolutional layer.

Amazon is well-known for personalization and recommendations, which help customers discover items they might otherwise not have found. In this update to their original paper, the authors discuss some of the changes as Amazon has grown.

We present a novel model called OCGAN for the classical problem of one-class novelty detection, where, given a set of examples from a particular class, the goal is to determine if a query example is from the same class. Our solution is based on learning latent representations of in-class examples using a de-noising auto-encoder network. The key contribution of our work is our proposal to explicitly constrain the latent space to exclusively represent the given class. In order to accomplish this goal, firstly, we force the latent space to have bounded support by introducing a tanh activation in the encoder's output layer. Secondly, using a discriminator in the latent space that is trained adversarially, we ensure that encoded representations of in-class examples resemble uniform random samples drawn from the same bounded space. Thirdly, using a second adversarial discriminator in the input space, we ensure all randomly drawn latent samples generate examples that look real. Finally, we introduce a gradient-descent based sampling technique that explores points in the latent space that generate potential out-of-class examples, which are fed back to the network to further train it to generate in-class examples from those points. The effectiveness of the proposed method is measured across four publicly available datasets using two one-class novelty detection protocols where we achieve state-of-the-art results.

We argue word embedding models are a useful tool for the study of culture using a historical analysis of shared understandings of social class as an empirical case. Word embeddings represent semantic relations between words as relationships between vectors in a high-dimensional space, specifying a relational model of meaning consistent with contemporary theories of culture. Dimensions induced by word differences ( rich – poor) in these spaces correspond to dimensions of cultural meaning, and the projection of words onto these dimensions reflects widely shared associations, which we validate with surveys. Analyzing text from millions of books published over 100 years, we show that the markers of class continuously shifted amidst the economic transformations of the twentieth century, yet the basic cultural dimensions of class remained remarkably stable. The notable exception is education, which became tightly linked to affluence independent of its association with cultivated taste.

The increased interest in analyzing and explaining gender inequalities in tech, media, and academia highlights the need for accurate inference methods to predict a person's gender from their name. Several such services exist that provide access to large databases of names, often enriched with information from social media profiles, culture-specific rules, and insights from sociolinguistics. We compare and benchmark five name-to-gender inference services by applying them to the classification of a test data set consisting of 7,076 manually labeled names. The compiled names are analyzed and characterized according to their geographical and cultural origin. We define a series of performance metrics to quantify various types of classification errors, and define a parameter tuning procedure to search for optimal values of the services' free parameters. Finally, we perform benchmarks of all services under study regarding several scenarios where a particular metric is to be optimized.

Assistive technology featuring artificial intelligence (AI) to support human decision-making has become ubiquitous. Assistive AI achieves accuracy comparable to or even surpassing that of human experts. However, often the adoption of assistive AI systems is limited by a lack of trust of humans into an AI’s prediction. This is why the AI research community has been focusing on rendering AI decisions more transparent by providing explanations of an AIs decision. To what extent these explanations really help to foster trust into an AI system remains an open question. In this paper, we report the results of a behavioural experiment in which subjects were able to draw on the support of an ML-based decision support tool for text classification. We experimentally varied the information subjects received and show that transparency can actually have a negative impact on trust. We discuss implications for decision makers employing assistive AI technology.

We explore the role of natural characteristics in determining the worldwide spatial distribution of economic activity, as proxied by lights at night, observed across 240,000 grid cells. A parsimonious set of 24 physical geography attributes explains 47% of worldwide variation and 35% of within-country variation in lights. We divide geographic characteristics into two groups, those primarily important for agriculture and those primarily important for trade, and confront a puzzle. In examining within-country variation in lights, among countries that developed early, agricultural variables incrementally explain over 6 times as much variation in lights as do trade variables, while among late developing countries the ratio is only about 1.5, even though the latter group is far more dependent on agriculture. Correspondingly, the marginal effects of agricultural variables as a group on lights are larger in absolute value, and those for trade smaller, for early developers than for late developers. We show that this apparent puzzle is explained by persistence and the differential timing of technological shocks in the two sets of countries. For early developers, structural transformation due to rising agricultural productivity began when transport costs were still high, so cities were localized in agricultural regions. When transport costs fell, these agglomerations persisted. In late-developing countries, transport costs fell before structural transformation. To exploit urban scale economies, manufacturing agglomerated in relatively few, often coastal, locations. Consistent with this explanation, countries that developed earlier are more spatially equal in their distribution of education and economic activity than late developers.

We study the problem of measuring group differences in choices when the dimensionality of the choice set is large. We show that standard approaches suffer from a severe finite‐sample bias, and we propose an estimator that applies recent advances in machine learning to address this bias. We apply this method to measure trends in the partisanship of congressional speech from 1873 to 2016, defining partisanship to be the ease with which an observer could infer a congressperson's party from a single utterance. Our estimates imply that partisanship is far greater in recent years than in the past, and that it increased sharply in the early 1990s after remaining low and relatively constant over the preceding century.

Training robust deep video representations has proven to be much more challenging than learning deep image representations. This is in part due to the enormous size of raw video streams and the high temporal redundancy; the true and interesting signal is often drowned in too much irrelevant data. Motivated by that the superfluous information can be reduced by up to two orders of magnitude by video compression (using H.264, HEVC, etc.), we propose to train a deep network directly on the compressed video. This representation has a higher information density, and we found the training to be easier. In addition, the signals in a compressed video provide free, albeit noisy, motion information. We propose novel techniques to use them effectively. Our approach is about 4.6 times faster than Res3D and 2.7 times faster than ResNet-152. On the task of action recognition, our approach outperforms all the other methods on the UCF-101, HMDB-51, and Charades dataset.

We are delighted to welcome you to the Eleventh ACM International Conference on Web Search and Data Mining (WSDM 2018) held in Los Angeles, California, USA, on February 5-9, 2018. Now in its eleventh year, WSDM has become a top tier conference in Web-inspired research relating to search and data mining. As in previous years, we continued observing a growth in number of submissions. The conference this year, with 514 valid submissions, maintained the major boost that was observed last year in Cambridge UK (with 505 submissions as compared to 386 in year 2016 in San Francisco). Interestingly, we saw an increase in 10% in abstract submissions with a record breaking 757 abstracts (as compared to 687 in 2017) but only 68% of those ended up as valid submissions. This is partly due to a relatively high number of invalid submissions that did not adhere to our new double-blind review policy. The 514 valid submissions originated from 41 countries, out of which 84 (as compared to 80 in 2017) were accepted for full paper publication in the proceedings, thus reaching an acceptance rate of 16.12% (as compared to 15.94% last year) and within the range of the last 11 years (with a min of 15.5% and a max of 22.3%.) Unfortunately, three papers were withdrawn/rejected after acceptance due to their violations of WSDM guidelines. The final 81 papers that will be published in the proceedings are from 23 countries, spanning four continents, making this a truly international forum. Oral presentation slots were allocated to all papers. Yet, in order to maintain the single track model that most attendees prefer, we followed the "spotlight" short presentation plus poster approach that was introduced in 2012. Out of the 81 accepted papers, 58 were assigned such a two-minute spotlight slot, while 23 were assigned a long twenty-minute talk slot. The type of slot was chosen by the Senior PC members and Program co-chairs, mostly based on whether the topic and the content of the paper were best suited for a large group presentation or for a more focused and interactive poster style of presentation. The double-blind flavor we adopted this year allowed the authors to indicate the source of their data set, or deployment environment (so as not to refer to "major commercial search engines" as has often been done in the past). However, we observed that several authors failed to indicate their conflict of interest (COI) adequately and we will make sure to enforce COI guidelines (through both awareness campaigns and hopefully automated tools) more rigorously now that the double-blind review policy has been adopted.

Recent work has achieved great success in utilizing global contextual information for semantic segmentation, including increasing the receptive field and aggregating pyramid feature representations. In this paper, we go beyond global context and explore the fine-grained representation using co-occurrent features by introducing Co-occurrent Feature Model, which predicts the distribution of co-occurrent features for a given target. To leverage the semantic context in the co-occurrent features, we build an Aggregated Co-occurrent Feature (ACF) Module by aggregating the probability of the co-occurrent feature with the co-occurrent context. ACF Module learns a fine-grained spatial invariant representation to capture co-occurrent context information across the scene. Our approach significantly improves the segmentation results using FCN and achieves superior performance 54.0% mIoU on Pascal Context, 87.2% mIoU on Pascal VOC 2012 and 44.89% mIoU on ADE20K datasets. The source code and complete system will be publicly available upon publication.

Deep learning based forecasting methods have become the methods of choice in many applications of time series prediction or forecasting often outperforming other approaches. Consequently, over the last years, these methods are now ubiquitous in large-scale industrial forecasting applications and have consistently ranked among the best entries in forecasting competitions (e.g., M4 and M5). This practical success has further increased the academic interest to understand and improve deep forecasting methods. In this article we provide an introduction and overview of the field: We present important building blocks for deep forecasting in some depth; using these building blocks, we then survey the breadth of the recent deep forecasting literature.

While recent progress on abstractive summarization has led to remarkably fluent summaries, factual errors in generated summaries still severely limit their use in practice. In this paper, we evaluate summaries produced by state-of-the-art models via crowdsourcing and show that such errors occur frequently, in particular with more abstractive models. We study whether textual entailment predictions can be used to detect such errors and if they can be reduced by reranking alternative predicted summaries. That leads to an interesting downstream application for entailment models. In our experiments, we find that outof-the-box entailment models trained on NLI datasets do not yet offer the desired performance for the downstream task and we therefore release our annotations as additional test data for future extrinsic evaluations of NLI.