Research Collection: Tools and Data to Advance the State of the Art

This post has been republished via RSS; it originally appeared at: Microsoft Research.

Editor’s Note: In the diverse and multifaceted world of research, individual contributions can add up to significant results over time. In this new series of posts, we’re connecting the dots to provide an overview of how researchers at Microsoft and their collaborators are working towards significant customer and societal outcomes that are broader than any single discipline. (Visit this collection to learn about the work Microsoft researchers are doing to advance responsible AI.) Here, we’ve curated a selection of the work Microsoft researchers are doing to advance the state of the art in tools and data research.

 

“This is a game changer for the big data community. Initiatives like Microsoft Research Open Data reduce barriers to data sharing and encourage reproducibility by leveraging the power of cloud computing” —Sam Madden, Professor, Massachusetts Institute of Technology

An open and collaborative research culture has always been central to research at Microsoft – as a researcher, it’s part of your job description to exchange ideas and share work among researchers, product engineers and developers at Microsoft, as well as the broader academic community.

Microsoft researchers and their collaborators have published tens of thousands of peer-reviewed papers since Microsoft Research’s founding in 1991. In the course of that research, they have created a wealth of tools and datasets that have been shared with the world to help advance the state of the art across all disciplines. They are available to be used by researchers and practitioners alike – not only in computer science and software development, but also in research across a wide variety of fields.

Below is a selection of those tools and datasets; visit our researcher tools page to see hundreds more, including code and data that can be used to reproduce the results of research papers.

Read our blog, listen to our podcast, and subscribe to our newsletter to stay up to date on all things research at Microsoft.

---

See sections on: MSR Open Data | AI and Machine Learning | Sound, Vision, Visualization and Interaction | Accessibility and Inclusion | Privacy and Security | Systems and Programming

---

MSR Open Data

Microsoft Research Open Data is a data repository that makes available datasets that researchers at Microsoft have created and published in conjunction with their research. Our goal is to provide a simple platform to Microsoft’s researchers and collaborators to share datasets and related research technologies and tools. The site has been designed to simplify access to these data sets, facilitate collaboration between researchers using cloud-based resources, and enable the reproducibility of research. Visit Microsoft Open Data for the latest datasets and related research assets.

AI and Machine Learning

AI Fairness Checklist

AI fairness checklists can help teams mitigate potential sources of unfairness in their products by providing organizational infrastructure for formalizing ad-hoc processes and empowering individual advocates of fairness. Through a series of semi-structured interviews and co-design workshops, we co-designed an AI fairness checklist and identified desiderata and concerns for AI fairness checklists in general. This checklist can help you prioritize fairness when developing AI systems.

AirSim

Microsoft AirSim (Aerial Informatics and Robotics Simulation) is an open-source robotics simulation platform. From ground vehicles, wheeled robotics, aerial drones, and even static IoT devices, AirSim can capture data for models without costly field operations. AirSim works as a plug-in to Epic Games’ Unreal Engine 4 editor, providing control over building environments and simulating difficult-to-reproduce, real-world events to capture meaningful data for AI models.

Datasheets for Datasets Template

Inspired by datasheets for electronic components, datasheets for datasets can be used to document characteristics of a dataset, including its motivation, composition, collection process, pre-processing, labeling, intended uses, distribution, and maintenance. Datasheets help dataset creators uncover potential sources of bias or hidden assumptions in their datasets. They help dataset consumers determine if a dataset meets their needs. Consider these questions to help prioritize transparency by creating a datasheet for your own dataset.

DeepSpeed / ZeRO

Part of Microsoft’s AI at Scale initiative, DeepSpeed is a PyTorch-compatible library that vastly improves large model training by improving scale, speed, cost and usability – unlocking the ability to train models with more than 100 billion parameters. One piece of the DeepSpeed library, ZeRO, is a new parallelized optimizer that greatly reduces the resources needed for model and data parallelism while massively increasing the number of parameters that can be trained.

DiCE: Diverse Counterfactual Explanations for Machine Learning

Explanations are critical for machine learning, especially as machine learning-based systems are being used to inform decisions in societally critical domains such as finance, healthcare, and education. Most explanation methods, however, approximate an ML model to create interpretable explanations. DiCE is a Python package that relies on counterfactuals to provide explanations that are always consistent with the ML model. Essentially, it provides “what-if” explanations to help explore how a model reacts to different inputs and what input changes are required to change the model’s output. Under the hood, DiCE optimizes the feasibility and diversity of generated explanations.

DoWhy: A Python Package for Causal Machine Learning

DoWhy is an end-to-end library for causal inference that allows data scientists to build causal models based on data and domain assumptions. Consider the problem of finding out why customers churn out of a service, deciding which drugs may work best for a patient, or recommending a financial or marketing strategy: all these questions require causal reasoning and cannot be addressed by predictive modeling alone. DoWhy systematizes the best practices for causal inference in a 4-step process: modeling the assumptions, identifying the correct estimand, estimating the effect, and finally checking robustness to violation of assumptions. While prior frameworks have focused on statistical estimation alone, DoWhy recognizes the criticality of formalizing domain assumptions and uses recent research in machine learning to validate assumptions as much possible.

Fairlearn

For those who wish to prioritize fairness in their AI systems, Microsoft researchers, in collaboration with Azure ML, have released Fairlearn—an open-source Python package that enables developers of AI systems to assess their systems’ fairness and mitigate any negative impacts for groups of people, such as those defined in terms of race, gender, age, or disability status. Fairlearn, which focuses specifically on harms of allocation or quality of service, draws on two papers by Microsoft researchers on incorporating quantitative fairness metrics into classification settings and regression settings, respectively.

Guidelines for Human-AI Interaction

Advances in AI are fundamentally changing how modern-day technologies behave and creating new challenges for developing intuitive, impactful, and responsible AI user experiences. The Guidelines for Human-AI Interaction synthesize and validate more than two decades of research and thinking on human-AI interaction into a set of 18 best practices for designers and developers of AI-based systems.

The Guidelines prescribe how an AI system should behave to set the right expectations, appropriately leverage a person’s social and environmental contexts, help a person recover when the AI is wrong, and carefully learn and improve over time. They are being used by interdisciplinary teams at Microsoft to create effective AI users experiences from the start when designing and planning as well as to evaluate existing experiences and identify opportunities for improvement.

InterpretML

Interpretability is essential for debugging machine learning models, detecting bias, gauging regulatory compliance, and using models in high-risk applications such as healthcare and finance. InterpretML is an open-source Python package for training interpretable machine learning models and explaining blackbox systems.

Historically, the most interpretable machine learning models were not very accurate, and the most accurate models were not very interpretable. Microsoft Research has developed an algorithm called the Explainable Boosting Machine (EBM), which has both high accuracy and interpretability. EBMs use modern machine learning techniques like bagging and gradient boosting to breathe new life into traditional GAMs (Generalized Additive Models). This makes them as accurate as random forests and gradient boosted trees, and also enhances their intelligibility and editability. EBMs allow users to learn more from their data, to deploy safer, more accurate models, and to detect and correct bias that might otherwise have remained hidden.

MMLSpark

Microsoft ML for Apache Spark expands the distributed computing framework Apache Spark in several new directions. MMLSpark adds several machine learning frameworks to the SparkML Ecosystem, including LightGBM, Vowpal Wabbit, OpenCV, Isolation Forest, and the Microsoft Cognitive Toolkit (CNTK) . These tools allow users to craft powerful and highly-scalable models that span multiple ML ecosystems.

MMLSpark also brings new networking capabilities to the Spark ecosystem. With the HTTP on Spark project, users can embed any web service into their SparkML models and use their Spark clusters for massive networking workflows. In this vein, MMLSpark provides easy to use SparkML transformers for a wide variety of Microsoft Cognitive Services. Finally, the Spark Serving project enables high throughput, sub-millisecond latency web services, backed by your Spark cluster.

Project Malmo

Project Malmo sets out to address the core research challenges of developing AI that learns to make sense of complex environments, learns how to interact with the world, learns transferable skills and applies them to new problems. Malmo is an AI experimentation platform built on top of Minecraft that is designed to support fundamental research in AI. It taps into Minecraft’s endless possibilities – from simple tasks like walking around looking for treasure, to complex ones like building a structure with a group of teammates. It consists of a mod for the Java version of Minecraft, and code that helps AI agents sense and act within that environment. These components can run on Windows, Linux or MacOS, and researchers can program their agents in any language they’re comfortable with.

TextWorld

Text-based games may seem primitive next to the beautifully rendered graphics of today, but to succeed at even the simplest such game, humans use a special set of skills. We seamlessly and simultaneously comprehend language descriptions, plan our next moves, visualize the environment, remember important information, and generalize past experiences to new situations. AI agents don’t yet possess these capabilities, but they are key to general intelligence. Microsoft TextWorld is an open-source, extensible platform that both generates and simulates text-based games. It is a popular testing ground for AI, used by researchers to train, evaluate, analyze, and improve reinforcement learning (RL) agents that learn from and interact with the world through language.

Sound, Vision, Visualization and Interaction

Phonetic Matching Library

Automatic Speech Recognition (ASR) systems tend to have trouble in noisy environments, when encountering unfamiliar accents, or when processing out-of-vocabulary words such as proper nouns and domain-specific words. Phonetic matching – comparing strings on a phoneme level rather than a character level – can help improve the accuracy and performance of these systems. Originally developed as a component of Maluuba’s natural language understanding platform, this library of phonetic string comparison utilities has been open sourced by Microsoft Research Montreal. It allows upstream systems to utilize personalized and contextual information that the Automatic Speech Recognition (ASR) systems may not have access to, in order to correct the ASR.

Project Rocket

Rocket is a Microsoft Research open-source platform that allows rapid construction of video pipelines for efficiently processing live video streams – with the goal of making it easy and affordable for anyone with a camera to benefit from video analytics. This highly extensible software stack empowers developers to build practical real-world applications for object detection and counting/alerting with cutting-edge machine learning algorithms. It features a hybrid edge-cloud video analytics pipeline (built on C# .NET Core), which allows TensorFlow, Darknet, ONNX and custom DNN model plug-in, GPU/FPGA acceleration, Docker containerization/Kubernetes orchestration, and interactive querying for after-the-fact analysis. Rocket’s pipelined architecture can be easily configured to execute over a distributed infrastructure, potentially spanning specialized edge hardware (e.g., Azure Stack Edge) and the cloud (e.g., Azure Machine Learning and Cognitive Services).

RocketBox Avatar Library

To empower research and academic institutions around the world to further investigate the relationship between people and their avatars and how it affects interactions with others in the virtual world, Microsoft is making the Microsoft Rocketbox Avatar Library—a collection of 115 avatars representing humans of different genders, races, and occupations—a publicly available resource for free research and academic use. Microsoft Rocketbox can be downloaded from GitHub. The release of the library coincides with last week’s celebration of the 2020 IEEE Conference on Virtual Reality and 3D User Interfaces (IEEE VR) and the presentation of our latest in avatar research, an active area here at Microsoft Research.

RoomAlive Toolkit

RoomAlive Toolkit is an open source SDK that transforms any room into an immersive, augmented, magical entertainment experience. RoomAlive Toolkit presents a unified, scalable approach for interactive projection mapping that dynamically adapts content to any room. Users can touch, shoot, stomp, dodge and steer projected content that seamlessly co-exists with their existing physical environment.

SandDance

SandDance is open-source software that enables you to more easily explore, identify, and communicate insights about your data. It provides easy to use, yet powerful, dynamic, data visualizations on large datasets, which enable pattern, trend, and outlier identification. These, in turn, provide better communication and decision making capabilities through an interface that can be embedded in your tool of choice: a custom visual in PowerBI, as extensions in Visual Studio Code and Azure Data Studio, as a stand-alone web-based application, or a react based control in a custom web application.

SandDance showcases Microsoft Research data visualization innovations and novel natural user interaction techniques and makes them available where you need them.

SandDance is available as a web-based application, as a custom visual in PowerBI, and as extensions in Visual Studio Code and Azure Data Studio.

Triton

Project Triton renders immersive acoustics that responds to the shape and materials of 3D environments in games and augmented/virtual reality (AR/VR). It is unique in modeling the true wave physics of sound, including diffraction, producing natural effects like smooth occlusion and scene-dependent reverberation. Detailed simulation is performed offline, yielding a lightweight runtime that scales from desktop to mobile phones. Incubated over a decade in Microsoft Research, Project Triton ships in major game titles like Gears of War, Sea of Thieves, and Borderlands 3. Plugin integrations for Unity and Unreal Engine are available as part of Project Acoustics.

Accessibility and Inclusion

Caption Crawler

Caption Crawler is a plug-in for the Edge and Chrome web browsers that provides additional information about images for screen reader users. Many images on the web lack captions (i.e., alt text). When a webpage loads, Caption Crawler identifies images that are missing captions and checks if it is captioned elsewhere on the web; if so, the discovered captions are made available to the screen reader.

Eyes First

Developed by Microsoft researchers, designers, and engineers, the Eyes First games help familiarize people with the basic skills of using eye tracking devices, which can be applied to other gaze-enabled assistive technologies. The games are designed to optimize the experiences for eye gaze input, but they are also mouse and touchscreen friendly and can be played by anyone with a compatible Windows 10 device.

The collection is powered by Windows 10 eye tracking APIs and can be used with or without Windows 10 Eye Control, a key accessibility feature that enables people with speech and mobility disabilities, including people living with ALS, to use their eyes to control an on-screen mouse, keyboard and text-to-speech experience.

The games are available for free from the Microsoft Store. The Microsoft Enable Experiences R&D team also mentors student groups on the development of additional eyes first apps and is actively expanding the Eyes First collection to include creativity experiences centered around inclusive music, art, and personal expression. This combination of research, human-centered design, empathy and creativity can help restore and redefine the experience of technology for people who are “locked in” or otherwise affected by severe speech and mobility impairment.

SeeingVR

SeeingVR is a research toolkit for making virtual reality more accessible to people with low vision – moderate to severe visual impairments that cannot be corrected fully by glasses. It includes a number of enhancements that provide visual and audio augmentations to existing VR experiences – including magnification, edge enhancement, brightness and contrast adjustment, text augmentation and text-to-speech. The toolkit also enables VR developers to add capabilities such as object descriptions and object labeling and highlighting. SeeingVR is designed to be used with the Unity VR development platform.

Soundscape

Microsoft Soundscape is a research project that explores the use of innovative audio-based technology to enable people, particularly those with blindness or low vision, to build a richer awareness of their surroundings, thus becoming more confident and empowered to get around. Soundscape provides information about your surroundings with synthesized binaural audio, creating the effect of 3D sound. Unlike step-by-step navigation apps, Soundscape uses audio cues to enrich ambient awareness and provide a new way to relate to the environment. It allows users to build a mental map and make personal route choices while being more comfortable within unfamiliar spaces.

Learn more about Microsoft Soundscape >

The post Research Collection: Tools and Data to Advance the State of the Art appeared first on Microsoft Research.