Showing papers by "Amazon.com published in 2015"

Sparse Convolutional Neural Networks

Abstract: Deep neural networks have achieved remarkable performance in both image classification and object detection problems, at the cost of a large number of parameters and computational complexity. In this work, we show how to reduce the redundancy in these parameters using a sparse decomposition. Maximum sparsity is obtained by exploiting both inter-channel and intra-channel redundancy, with a fine-tuning step that minimize the recognition loss caused by maximizing sparsity. This procedure zeros out more than 90% of parameters, with a drop of accuracy that is less than 1% on the ILSVRC2012 dataset. We also propose an efficient sparse matrix multiplication algorithm on CPU for Sparse Convolutional Neural Networks (SCNN) models. Our CPU implementation demonstrates much higher efficiency than the off-the-shelf sparse matrix libraries, with a significant speedup realized over the original dense network. In addition, we apply the SCNN model to the object detection problem, in conjunction with a cascade model and sparse fully connected layers, to achieve significant speedups.

Long-Time-Step Molecular Dynamics through Hydrogen Mass Repartitioning.

Abstract: Previous studies have shown that the method of hydrogen mass repartitioning (HMR) is a potentially useful tool for accelerating molecular dynamics (MD) simulations. By repartitioning the mass of heavy atoms into the bonded hydrogen atoms, it is possible to slow the highest-frequency motions of the macromolecule under study, thus allowing the time step of the simulation to be increased by up to a factor of 2. In this communication, we investigate further how this mass repartitioning allows the simulation time step to be increased in a stable fashion without significantly increasing discretization error. To this end, we ran a set of simulations with different time steps and mass distributions on a three-residue peptide to get a comprehensive view of the effect of mass repartitioning and time step increase on a system whose accessible phase space is fully explored in a relatively short amount of time. We next studied a 129-residue protein, hen egg white lysozyme (HEWL), to verify that the observed behavior extends to a larger, more-realistic, system. Results for the protein include structural comparisons from MD trajectories, as well as comparisons of pKa calculations via constant-pH MD. We also calculated a potential of mean force (PMF) of a dihedral rotation for the MTS [(1-oxyl-2,2,5,5-tetramethyl-pyrroline-3-methyl)methanethiosulfonate] spin label via umbrella sampling with a set of regular MD trajectories, as well as a set of mass-repartitioned trajectories with a time step of 4 fs. Since no significant difference in kinetics or thermodynamics is observed by the use of fast HMR trajectories, further evidence is provided that long-time-step HMR MD simulations are a viable tool for accelerating MD simulations for molecules of biochemical interest.

Long-term decline of the Amazon carbon sink

Abstract: Atmospheric carbon dioxide records indicate that the land surface has acted as a strong global carbon sink over recent decades, with a substantial fraction of this sink probably located in the tropics, particularly in the Amazon. Nevertheless, it is unclear how the terrestrial carbon sink will evolve as climate and atmospheric composition continue to change. Here we analyse the historical evolution of the biomass dynamics of the Amazon rainforest over three decades using a distributed network of 321 plots. While this analysis confirms that Amazon forests have acted as a long-term net biomass sink, we find a long-term decreasing trend of carbon accumulation. Rates of net increase in above-ground biomass declined by one-third during the past decade compared to the 1990s. This is a consequence of growth rate increases levelling off recently, while biomass mortality persistently increased throughout, leading to a shortening of carbon residence times. Potential drivers for the mortality increase include greater climate variability, and feedbacks of faster growth on mortality, resulting in shortened tree longevity. The observed decline of the Amazon sink diverges markedly from the recent increase in terrestrial carbon uptake at the global scale, and is contrary to expectations based on models.

Scalable distributed DNN training using commodity GPU cloud computing.

Abstract: We introduce a new method for scaling up distributed Stochastic Gradient Descent (SGD) training of Deep Neural Networks (DNN). The method solves the well-known communication bottleneck problem that arises for data-parallel SGD because compute nodes frequently need to synchronize a replica of the model. We solve it by purposefully controlling the rate of weight-update per individual weight, which is in contrast to the uniform update-rate customarily imposed by the size of a mini-batch. It is shown empirically that the method can reduce the amount of communication by three orders of magnitude while training a typical DNN for acoustic modelling. This reduction in communication bandwidth enables efficient scaling to more parallel GPU nodes than any other method that we are aware of, and it can be achieved with neither loss in convergence rate nor accuracy in the resulting DNN. Furthermore, the training can be performed on commodity cloud infrastructure and networking.

Brazil's Soy Moratorium

Abstract: Brazil's Soy Moratorium (SoyM) was the first voluntary zero-deforestation agreement implemented in the tropics and set the stage for supply-chain governance of other commodities, such as beef and palm oil [supplementary material (SM)]. In response to pressure from retailers and nongovernmental organizations (NGOs), major soybean traders signed the SoyM, agreeing not to purchase soy grown on lands deforested after July 2006 in the Brazilian Amazon. The soy industry recently extended the SoyM to May 2016, by which time they assert that Brazil's environmental governance, such as the increased enforcement and national implementation of the Rural Environmental Registry of private properties (Portuguese acronym CAR) mandated by the Forest Code (FC) ( 1 ), will be robust enough to justify ending the agreement ( 2 ). We argue that a longer-term commitment is needed to help maintain deforestation-free soy supply chains, as full compliance and enforcement of these regulations is likely years away. Ending the SoyM prematurely would risk a return to deforestation for soy expansion at a time when companies are committing to zero-deforestation supply chains ( 3 ).

Scikit-learn: Machine Learning Without Learning the Machinery

Abstract: Machine learning is a pervasive development at the intersection of statistics and computer science. While it can benefit many data-related applications, the technical nature of the research literature and the corresponding algorithms slows down its adoption. Scikit-learn is an open-source software project that aims at making machine learning accessible to all, whether it be in academia or in industry. It benefits from the general-purpose Python language, which is both broadly adopted in the scientific world, and supported by a thriving ecosystem of contributors. Here we give a quick introduction to scikit-learn as well as to machine-learning basics.

How Amazon web services uses formal methods

Abstract: Engineers use TLA+ to prevent serious but subtle bugs from reaching production.

A robust and efficient video representation for action recognition

Abstract: This paper introduces a state-of-the-art video representation and applies it to efficient action recognition and detection. We first propose to improve the popular dense trajectory features by explicit camera motion estimation. More specifically, we extract feature point matches between frames using SURF descriptors and dense optical flow. The matches are used to estimate a homography with RANSAC. To improve the robustness of homography estimation, a human detector is employed to remove outlier matches from the human body as human motion is not constrained by the camera. Trajectories consistent with the homography are considered as due to camera motion, and thus removed. We also use the homography to cancel out camera motion from the optical flow. This results in significant improvement on motion-based HOF and MBH descriptors. We further explore the recent Fisher vector as an alternative feature encoding approach to the standard bag-of-words histogram, and consider different ways to include spatial layout information in these encodings. We present a large and varied set of evaluations, considering (i) classification of short basic actions on six datasets, (ii) localization of such actions in feature-length movies, and (iii) large-scale recognition of complex events. We find that our improved trajectory features significantly outperform previous dense trajectories, and that Fisher vectors are superior to bag-of-words encodings for video recognition tasks. In all three tasks, we show substantial improvements over the state-of-the-art results.

Using virtual networking devices to connect managed computer networks

Abstract: Techniques are described for providing managed virtual computer networks whose configured logical network topology may have one or more virtual networking devices, such as by a network-accessible configurable network service, with corresponding networking functionality provided for communications between multiple computing nodes of a virtual computer network by emulating functionality that would be provided by the networking devices if they were physically present. The networking functionality provided for a managed computer network may include supporting a connection between that managed computer network and other managed computer networks, such as via a provided virtual peering router to which each of the managed computer networks may connect, with the functionality of the virtual peering router being emulated by modules of the configurable network service without physically providing the virtual peering router, including to manage data communications between computing nodes of the inter-connected managed computer networks in accordance with client-specified configuration information.

Using virtual networking devices to manage routing communications between connected computer networks

Abstract: Techniques are described for providing managed virtual computer networks whose configured logical network topology may have one or more virtual networking devices, such as by a network-accessible configurable network service, with corresponding networking functionality provided for communications between multiple computing nodes of a virtual computer network by emulating functionality that would be provided by the networking devices if they were physically present. The networking functionality provided for a managed computer network may include supporting a connection between that managed computer network and one or more other managed computer networks, such as via a provided virtual peering router to which each of the managed computer networks may connect, with the functionality of the virtual peering router being emulated by modules of the configurable network service without physically providing the virtual peering router, including to manage routing communications between the inter-connected managed computer networks in accordance with client-specified configuration information.

Interactive interfaces for machine learning model evaluations

Abstract: A first data set corresponding to an evaluation run of a model is generated at a machine learning service for display via an interactive interface. The data set includes a prediction quality metric. A target value of an interpretation threshold associated with the model is determined based on a detection of a particular client's interaction with the interface. An indication of a change to the prediction quality metric that results from the selection of the target value may be initiated.

System for managing and scheduling containers

Abstract: A system and method for a container service that obtains a software image of a software container that has been configured to be executed within a computer system instance registered to a cluster by one or more processors. The container service is configured to receive a request to launch the software image in accordance with a task definition, wherein the task definition specifies an allocation of resources for the software container. The container service may then determine, according to a placement scheme, a subset of a set of container instances registered to the cluster in which to launch the software image in accordance with the task definition. Upon determining the subset of the set of container instances, the container service may launch the software image as one or more running software containers in the set of container instances in accordance with the task definition.

Deployment and adjustment of airborne unmanned aerial vehicles

Abstract: This disclosure describes a power unmanned aerial vehicle (UAV) that may generate a current from a magnetic field of an overhead power line. In various implementations, while the power UAV is flying, the power UAV may receive another UAV at a platform. A control element of the power UAV may generate signals to cause the power UAV to fly to a location of a conductor of the power line. The control element may also determine a position of the secondary coil with respect to the power line and generate control signals to adjust the position of the secondary coil based on the determined position of the secondary coil, a determined safety distance, and/or a determined threshold distance for efficient current generation. A shielding substrate may also be provided to shield electronics of the power UAV or other UAVs from magnetic fields.

Device Selection for Providing a Response

Abstract: A system may use multiple speech interface devices to interact with a user by speech. All or a portion of the speech interface devices may detect a user utterance and may initiate speech processing to determine a meaning or intent of the utterance. Within the speech processing, arbitration is employed to select one of the multiple speech interface devices to respond to the user utterance. Arbitration may be based in part on metadata that directly or indirectly indicates the proximity of the user to the devices, and the device that is deemed to be nearest the user may be selected to respond to the user utterance.

Better Document-level Sentiment Analysis from RST Discourse Parsing

Abstract: Discourse structure is the hidden link between surface features and document-level properties, such as sentiment polarity. We show that the discourse analyses produced by Rhetorical Structure Theory (RST) parsers can improve document-level sentiment analysis, via composition of local information up the discourse tree. First, we show that reweighting discourse units according to their position in a dependency representation of the rhetorical structure can yield substantial improvements on lexicon-based sentiment analysis. Next, we present a recursive neural network over the RST structure, which offers significant improvements over classificationbased methods.

Dynamic Service Placement for Mobile Micro-Clouds with Predicted Future Costs

Abstract: Mobile micro-clouds are promising for enabling performance-critical cloud applications. However, one challenge therein is the dynamics at the network edge. In this paper, we study how to place service instances to cope with these dynamics, where multiple users and service instances coexist in the system. Our goal is to find the optimal placement (configuration) of instances to minimize the average cost over time, leveraging the ability of predicting future cost parameters with known accuracy. We first propose an offline algorithm that solves for the optimal configuration in a specific look-ahead time-window. Then, we propose an online approximation algorithm with polynomial time-complexity to find the placement in real-time whenever an instance arrives. We analytically show that the online algorithm is $O(1)$-competitive for a broad family of cost functions. Afterwards, the impact of prediction errors is considered and a method for finding the optimal look-ahead window size is proposed, which minimizes an upper bound of the average actual cost. The effectiveness of the proposed approach is evaluated by simulations with both synthetic and real-world (San Francisco taxi) user-mobility traces. The theoretical methodology used in this paper can potentially be applied to a larger class of dynamic resource allocation problems.

Electronic communication with secure screen sharing of sensitive information

Abstract: A customer support application provides screen sharing of the user's computing device with a remote customer support agent, thereby enabling the customer support agent to view the content displayed on the user's device. Sensitive information that is displayed on a user's computing device is obfuscated from the computing device of the remote customer support agent, and a notification of that obfuscation is displayed on the user's computing device. Information can be determined to be sensitive based on a sensitive indicator tag or a heuristic.

Landing of unmanned aerial vehicles on transportation vehicles for transport

Abstract: Unmanned aerial vehicles (“UAVs”) which fly to destinations (e.g., for delivering items) may land on transportation vehicles (e.g., delivery trucks, etc.) for temporary transport. An agreement with the owner of the transportation vehicles (e.g., a shipping carrier) may be made for obtaining consent and determining compensation for landings, and the associated transportation vehicles that are available for landings may be identified by markers on the roof or other identification techniques. The routes of the transportation vehicles may be known and utilized to determine locations where UAVs will land on and take off from the transportation vehicles, and in cases of emergencies (e.g., due to low batteries, mechanical issues, etc.) the UAVs may land on the transportation vehicles for later retrieval.

Estimating the global conservation status of more than 15,000 Amazonian tree species

Abstract: Estimates of extinction risk for Amazonian plant and animal species are rare and not often incorporated into land-use policy and conservation planning. We overlay spatial distribution models with historical and projected deforestation to show that at least 36% and up to 57% of all Amazonian tree species are likely to qualify as globally threatened under International Union for Conservation of Nature (IUCN) Red List criteria. If confirmed, these results would increase the number of threatened plant species on Earth by 22%. We show that the trends observed in Amazonia apply to trees throughout the tropics, and we predict that most of the world’s >40,000 tropical tree species now qualify as globally threatened. A gap analysis suggests that existing Amazonian protected areas and indigenous territories will protect viable populations of most threatened species if these areas suffer no further degradation, highlighting the key roles that protected areas, indigenous peoples, and improved governance can play in preventing large-scale extinctions in the tropics in this century.

Arbitration between voice-enabled devices

Abstract: Architectures and techniques for selecting a voice-enabled device to handle audio input that is detected by multiple voice-enabled devices are described herein. In some instances, multiple voice-enabled devices may detect audio input from a user at substantially the same time, due to the voice-enabled devices being located within proximity to the user. The architectures and techniques may analyze a variety of audio signal metric values for the voice-enabled devices to designate a voice-enabled device to handle the audio input.

Route Planning in Transportation Networks

Abstract: We survey recent advances in algorithms for route planning in transportation networks. For road networks, we show that one can compute driving directions in milliseconds or less even at continental scale. A variety of techniques provide different trade-offs between preprocessing effort, space requirements, and query time. Some algorithms can answer queries in a fraction of a microsecond, while others can deal efficiently with real-time traffic. Journey planning on public transportation systems, although conceptually similar, is a significantly harder problem due to its inherent time-dependent and multicriteria nature. Although exact algorithms are fast enough for interactive queries on metropolitan transit systems, dealing with continent-sized instances requires simplifications or heavy preprocessing. The multimodal route planning problem, which seeks journeys combining schedule-based transportation (buses, trains) with unrestricted modes (walking, driving), is even harder, relying on approximate solutions even for metropolitan inputs.

Robot-Centric Activity Prediction from First-Person Videos: What Will They Do to Me?

Abstract: In this paper, we present a core technology to enable robot recognition of human activities during human-robot interactions. In particular, we propose a methodology for early recognition of activities from robot-centric videos (i.e., first-person videos) obtained from a robot's viewpoint during its interaction with humans. Early recognition, which is also known as activity prediction, is an ability to infer an ongoing activity at its early stage. We present an algorithm to recognize human activities targeting the camera from streaming videos, enabling the robot to predict intended activities of the interacting person as early as possible and take fast reactions to such activities (e.g., avoiding harmful events targeting itself before they actually occur). We introduce the novel concept of'onset' that efficiently summarizes pre-activity observations, and design a recognition approach to consider event history in addition to visual features from first-person videos. We propose to represent an onset using a cascade histogram of time series gradients, and we describe a novel algorithmic setup to take advantage of such onset for early recognition of activities. The experimental results clearly illustrate that the proposed concept of onset enables better/earlier recognition of human activities from first-person videos collected with a robot. Categories and Subject Descriptors I.2.10 [Artificial Intelligence]: Vision and Scene Understanding–video analysis; I.4.8 [Image Processing and Computer Vision]: Scene Analysis-motion; I.2.9 [Artificial Intelligence]: Robotics–sensors

Amazon Redshift and the Case for Simpler Data Warehouses

Abstract: Amazon Redshift is a fast, fully managed, petabyte-scale data warehouse solution that makes it simple and cost-effective to efficiently analyze large volumes of data using existing business intelligence tools. Since launching in February 2013, it has been Amazon Web Service's (AWS) fastest growing service, with many thousands of customers and many petabytes of data under management. Amazon Redshift's pace of adoption has been a surprise to many participants in the data warehousing community. While Amazon Redshift was priced disruptively at launch, available for as little as $1000/TB/year, there are many open-source data warehousing technologies and many commercial data warehousing engines that provide free editions for development or under some usage limit. While Amazon Redshift provides a modern MPP, columnar, scale-out architecture, so too do many other data warehousing engines. And, while Amazon Redshift is available in the AWS cloud, one can build data warehouses using EC2 instances and the database engine of one's choice with either local or network-attached storage. In this paper, we discuss an oft-overlooked differentiating characteristic of Amazon Redshift -- simplicity. Our goal with Amazon Redshift was not to compete with other data warehousing engines, but to compete with non-consumption. We believe the vast majority of data is collected but not analyzed. We believe, while most database vendors target larger enterprises, there is little correlation in today's economy between data set size and company size. And, we believe the models used to procure and consume analytics technology need to support experimentation and evaluation. Amazon Redshift was designed to bring data warehousing to a mass market by making it easy to buy, easy to tune and easy to manage while also being fast and cost-effective.

Automatic scaling of resource instance groups within compute clusters

Abstract: A service provider may apply customer-selected or customer-defined auto-scaling policies to a cluster of resources (e.g., virtualized computing resource instances or storage resource instances in a MapReduce cluster). Different policies may be applied to different subsets of cluster resources (e.g., different instance groups containing nodes of different types or having different roles). Each policy may define an expression to be evaluated during execution of a distributed application, a scaling action to take if the expression evaluates true, and an amount by which capacity should be increased or decreased. The expression may be dependent on metrics emitted by the application, cluster, or resource instances by default, metrics defined by the client and emitted by the application, or metrics created through aggregation. Metric collection, aggregation and rules evaluation may be performed by a separate service or by cluster components. An API may support auto-scaling policy definition.

Application focus in speech-based systems

Abstract: A speech-based system includes an audio device in a user premises and a network-based service that supports use of the audio device by multiple applications. The audio device may be directed to play audio content such as music, audio books, etc. The audio device may also be directed to interact with a user through speech. The network-based service monitors event messages received from the audio device to determine which of the multiple applications currently has speech focus. When receiving speech from a user, the service first offers the corresponding meaning to the application, if any, that currently has primary speech focus. If there is no application that currently has primary speech focus, or if the application having primary speech focus is not able to respond to the meaning, the service then offers the user meaning to the application that currently has secondary speech focus.

Context-sensitive learning for enhanced audiovisual emotion classification (Extended abstract)

Abstract: Human emotional expression tends to evolve in a structured manner in the sense that certain emotional evolution patterns, i.e., anger to anger, are more probable than others, e.g., anger to happiness. Furthermore the perception of an emotional display can be affected by recent emotional displays. Therefore, the emotional content of past and future observations could offer relevant temporal context when classifying the emotional content of an observation. In this work, we focus on audio-visual recognition of the emotional content of improvised emotional interactions at the utterance level. We examine context-sensitive schemes for emotion recognition within a multimodal, hierarchical approach: bidirectional Long Short-Term Memory (BLSTM) neural networks, hierarchical Hidden Markov Model classifiers (HMMs) and hybrid HMM/BLSTM classifiers are considered for modeling emotion evolution within an utterance and between utterances over the course of a dialog. Overall, our experimental results indicate that incorporating long-term temporal context is beneficial for emotion recognition systems that encounter a variety of emotional manifestations.

Methods and devices for selectively ignoring captured audio data

Abstract: Systems and methods for selectively ignoring an occurrence of a wakeword within audio input data is provided herein. In some embodiments, a wakeword may be detected to have been uttered by an individual within a modified time window, which may account for hardware delays and echoing offsets. The detected wakeword that occurs during this modified time window may, in some embodiments, correspond to a word included within audio that is outputted by a voice activated electronic device. This may cause the voice activated electronic device to activate itself, stopping the audio from being outputted. By identifying when these occurrences of the wakeword within outputted audio are going to happen, the voice activated electronic device may selectively determine when to ignore the wakeword, and furthermore, when not to ignore the wakeword.