Track Session Program - All times EST

DAY 1
Monday, December 14

11:00 AM - 12:30 PM: Track Session 1

Enabling New Applications

Talk Title: Holistic Hardware Optimizations for Machine Learning Acceleration

Abstract: Machine learning systems are being widely deployed across billions of edge devices and datacenter across the world. At the same time, in the absence of Moore’s Law and Dennard scaling, we rely on building vertically integrated systems with domain-specific accelerators to improve system performance and efficiency. In this talk, I will discuss challenges and opportunities in designing domain-specific architectures for machine learning applications and some of our recent work on developing holistic hardware optimizations for machine learning applications to improve scalability, performance, and programmability.

Talk Title: MLCommons: Accelerating Machine Learning Innovation

Abstract: The new MLCommons organization aims to accelerate machine learning innovation to benefit everyone. Machine learning has tremendous potential to save lives in areas like healthcare and automotive safety and to improve information access and understanding through technologies like voice interfaces, automatic translation, and natural language processing. However, machine learning is completely unlike conventional software and requires a whole new set of techniques analogous to the breakthroughs in precision measurement, raw materials, and manufacturing that drove the industrial revolution.MLCommons aims to answer the needs of the nascent machine learning industry through open, collaborative engineering in three areas: benchmarks, public datasets, and best practices. This talk will provide a technical overview of MLCommons activities including the MLPerf benchmark suite, the People's Speech dataset, and the MLCube model-sharing best practice.

Talk Title: Similarity Search at Massive Scale with Faiss

Abstract: Similarity search is a key part of the modern machine learning toolbox. Based upon a primitive of finding nearest neighbors of high-dimensional vectors, it can be applied at scale to translate between languages without depending upon parallel texts or dictionaries, search for similar images among billions of candidates in microseconds, or significantly accelerate neural network inference. I will present our work on Faiss, Facebook’s open-sourced similarity search library, and show how it is used to solve these and related challenges in research and production.

2:30 PM - 4:00 PM: Track Session 3

Great Free Tools

Talk Title: OpenDP, an Open-Source Suite of Tools for Deploying Differential Privacy

Abstract: Since it was introduced in 2006, differential privacy (DP) has become accepted as a gold standard for ensuring that individual-level information is not leaked through statistical analyses or machine learning on sensitive datasets. OpenDP comes at a time when computation and methodological advances, together with a growing need to analyze sensitive data while protecting privacy, are moving DP from theory to practice. In fact, in recent years, DP has seen large-scale deployments by Google, Apple, Microsoft, and the US Census Bureau, all organizations with the resources and expertise to implement their own custom DP systems. What does OpenDP add? OpenDP brings together an open-source community that can contribute to and ensure the trustworthiness of a DP library and an accompanying suite of tools to generate DP statistical releases. This talk describes our efforts to develop the initial OpenDP components and build its community.

Talk Title: Fairness in Machine Learning with Fairlearn

Abstract: Artificial Intelligence is being used in more and more fields, and AI systems are being used to make consequential decisions. The range is broad, encompassing decisions about which resumes to forward to hiring managers, deciding on creditworthiness for loans, determining which defendants are offered bail, and many more besides. In this talk, we will discuss some of the ways in which AI systems can behave unfairly, and what can be done to mitigate these behaviors. The session will demonstrate the use of the Fairlearn package (an open source toolkit for assessing AI systems’ fairness and mitigating their fairness issues) which incorporates metrics, dashboards, and algorithms for measuring and mitigating disparities in AI models.

Talk Title: Pangeo - Open Source Tools for Big Data (Climate) Science

Abstract: Modern scientific discoveries are increasingly dependent on big datasets (many petabytes) and will further increase in size in the future. This makes the traditional 'download and analyze on your laptop' model of data-analysis with traditional tools unfeasible.

Pangeo is a community of scientists and developers that aims to provide an ecosystem of fast, flexible, and scalable open-source tools to enable cutting edge science on big datasets. 

In this presentation, I will give an overview of the motivation, guiding principles at the core of pangeo. The highlighted tools include Xarray (high level, label aware operations on ND-arrays), Dask (flexible, general-purpose parallel computing framework), and Zarr (open source library for storage of chunked, compressed ND-arrays). I will conclude with an interactive demo using climate data in the cloud.

Details to be shared in session

Participating Speakers & Track Chairs

Research Submissions

Submission Deadline: October 15, 2020
Acceptance Decision: November 15, 2020

The Computing Systems track will accept 3 types of submissions. All submissions are limited to 2 pages, maximum:

A research highlight (2 pages) highlights recent work and contextualizes it for the broader MLSE audience.

An extended abstract (2 pages) describes unpublished original research.  These abstracts are non-archival.

A position paper (2 pages) argues a viewpoint, supported by published and original research.