Amazon.com

About: Amazon.com is a company organization based out in Seattle, Washington, United States. It is known for research contribution in the topics: Computer science & Service (business). The organization has 13363 authors who have published 17317 publications receiving 266589 citations.

LatticeRnn: Recurrent Neural Networks Over Lattices.

Abstract: We present a new model called LATTICERNN, which generalizes recurrent neural networks (RNNs) to process weighted lattices as input, instead of sequences. A LATTICERNN can encode the complete structure of a lattice into a dense representation, which makes it suitable to a variety of problems, including rescoring, classifying, parsing, or translating lattices using deep neural networks (DNNs). In this paper, we use LATTICERNNs for a classification task: each lattice represents the output from an automatic speech recognition (ASR) component of a spoken language understanding (SLU) system, and we classify the intent of the spoken utterance based on the lattice embedding computed by a LATTICERNN. We show that making decisions based on the full ASR output lattice, as opposed to 1-best or n-best hypotheses, makes SLU systems more robust to ASR errors. Our experiments yield improvements of 13% over a baseline RNN system trained on transcriptions and 10% over an nbest list rescoring system for intent classification.

Using gaze determination with device input

Abstract: A computing device, in a locked operational state, captures image information of a user which is analyzed to determine the direction of the users gaze. When the users gaze is determined to be substantially in the direction of the device, a predetermined input from the user, such as a tap or a voice command, will provide the user with access to at least some functionality of the device that was previously unavailable. If, however, the computing device detects what appears to be the predetermined input, but the users gaze direction is not in the direction of the device, the computing device will remain in the locked operational state. Therefore, in accordance with various embodiments, gaze determination is utilized as an indication that the user intends to unlock at least some additional functionality of the computing device.

Why Amazon Chose TLA

Abstract: Since 2011, engineers at Amazon have been using TLA+ to help solve difficult design problems in critical systems. This paper describes the reasons why we chose TLA+ instead of other methods, and areas in which we would welcome further progress.

Pan-tropical prediction of forest structure from the largest trees

Abstract: Aim Large tropical trees form the interface between ground and airborne observations, offering a unique opportunity to capture forest properties remotely and to investigate their variations on broad scales. However, despite rapid development of metrics to characterize the forest canopy from remotely sensed data, a gap remains between aerial and field inventories. To close this gap, we propose a new pan-tropical model to predict plot-level forest structure properties and biomass from only the largest trees. Location Time period Pan-tropical. Early 21st century. Major taxa studied Methods Woody plants. Using a dataset of 867 plots distributed among 118 sites across the tropics, we tested the prediction of the quadratic mean diameter, basal area, Lorey's height, community wood density and aboveground biomass (AGB) from the ith largest trees. Results Main conclusions Measuring the largest trees in tropical forests enables unbiased predictions of plot- and site-level forest structure. The 20 largest trees per hectare predicted quadratic mean diameter, basal area, Lorey's height, community wood density and AGB with 12, 16, 4, 4 and 17.7% of relative error, respectively. Most of the remaining error in biomass prediction is driven by differences in the proportion of total biomass held in medium-sized trees (50-70 cm diameter at breast height), which shows some continental dependency, with American tropical forests presenting the highest proportion of total biomass in these intermediate-diameter classes relative to other continents. Our approach provides new information on tropical forest structure and can be used to generate accurate field estimates of tropical forest carbon stocks to support the calibration and validation of current and forthcoming space missions. It will reduce the cost of field inventories and contribute to scientific understanding of tropical forest ecosystems and response to climate change.

Rethinking Zero-Shot Video Classification: End-to-End Training for Realistic Applications

Abstract: Trained on large datasets, deep learning (DL) can accurately classify videos into hundreds of diverse classes. However, video data is expensive to annotate. Zero-shot learning (ZSL) proposes one solution to this problem. ZSL trains a model once, and generalizes to new tasks whose classes are not present in the training dataset. We propose the first end-to-end algorithm for ZSL in video classification. Our training procedure builds on insights from recent video classification literature and uses a trainable 3D CNN to learn the visual features. This is in contrast to previous video ZSL methods, which use pretrained feature extractors. We also extend the current benchmarking paradigm: Previous techniques aim to make the test task unknown at training time but fall short of this goal. We encourage domain shift across training and test data and disallow tailoring a ZSL model to a specific test dataset. We outperform the state-of-the-art by a wide margin. Our code, evaluation procedure and model weights are available online github.com/bbrattoli/ZeroShotVideoClassification.