We are undoubtedly living in the Age of Data, where data is power and currency. The massive and complex datasets being collected via the Internet and scientific domains, including genomics, astronomy, physics, and finance, hold the promise for many transformational applications, and machine learning techniques are seemingly poised to deliver on this promise. However, concerns over scalability and ease-of-use present major roadblocks to the wider adoption of these statistical methods. For instance, modern genomic sequencing technologies can rapidly generate hundreds of gigabytes of raw data for a single individual, yet the cost to analyze this data is emerging as a major obstacle in eventually achieving the goal of personalized medicine. As a second example, measuring customer activity is no longer a practice limited to pioneering companies like Google and Netflix, yet many organizations lack the statistical expertise required to make actionable decisions from their data.

Dr. Ameet Talwalkar, Assistant Professor of Computer Science at the University of California, Los Angeles, addresses scalability and ease-of-use issues like these, and aims to ultimately connect advances in machine learning back to the real problems in science and technology. Having led the initial development of MLlib, a leading distributed machine learning library that is part of the Apache Spark project, and having co-authored a graduate level textbook about machine learning called Foundations of Machine Learning, Dr. Talwalkar is fully equipped with both the knowledge and the expertise to tackle these challenges. He continues to educate researchers and practitioners through his massive open online course (MOOC) on edX as well as other sources.

Current research projects include:

• Scalability: Data analyses suitable for modest-sized datasets are often entirely infeasible for the terabyte and petabyte datasets that are quickly becoming the norm. Over the past decade, Dr. Talwalkar has devised and analyzed scalable learning algorithms that leverage distributed (e.g., cloud-based) computing architectures, focusing on a variety of problems including collaborative filtering for recommender systems, estimating the uncertainty or quality of learning models, and facilitating the processing of genomic sequencing data. He often employs a divide-and-conquer strategy, whereby he builds upon existing base algorithms that have proven their value at smaller scales but executes them on subsamples of the data, thus yielding efficient approximation algorithms. Currently, he is working on scalable approximation methods for learning algorithms such as deep learning and random forests, two state-of-the-art methods for classification and regression tasks with a wide range of applications including image classification, protein function prediction and speech recognition.

• Ease-of-use: Successfully deploying data processing pipelines is currently a highly manual process, and is fraught with pitfalls for domain experts lacking strong statistical backgrounds. The process of preprocessing input data, selecting the appropriate learning model, and tuning the various knobs of these models can be an ad-hoc and a time-consuming task. To address these issues, Dr. Talwalkar is developing principled tools to autotune components of typical machine learning pipelines. The long-term vision is to build a system called MLbase that leverages these tools to automatically construct machine learning pipelines. Through abstractions, MLbase will enable end users to issue high-level queries and reason about predicted attributes, without the burden of understanding the low-level details.