Machines That Teach Themselves

RAJAN BHATTACHARYYA
HRL Laboratories

The human race has been using tools for more than 2.5 million years, and building machines for just more than 2,000 years. Over the past 200 years, humans developed machines to do physical work during the Industrial Age, and in the past 50 years innovations in technology areas such as electronics and computer science spawned the Digital Age.

Until now, machines were designed by hand to perform specialized functions in a highly efficient way using engineering principles. In the Information Age, data volume is increasing by 40 percent annually and streaming at faster rates each year. Moreover, this exponential growth is dominated by an acceleration in unstructured data due to the variety of sources, which include documents, video, audio, and embedded sensors. Finally, high dimensionality and uncertainty in data require new computational methods to extract latent patterns and semantics. Taken together, these challenges necessitate a new way to build machines to make Information Age data useful.

In this session the speakers explored machines that process information into useful output in a variety of applications but that are optimized in a very different way: by learning their own models. Emma Brunskill (Stanford University) opened the session with a presentation on how interactive machine learning can be applied to self-optimizing tutoring systems in classrooms.¹ Her work advances the paradigm of reinforcement learning, an important pillar in building machines that teach themselves. Suchi Saria (Johns Hopkins University) focused on machine systems that utilize highly heterogeneous data, ranging from sensor streams and genomic data to unstructured data, such as text, to perform inference.¹ She

___________________

¹ Paper not included in this volume.

explained how she applies a variety of machine learning methods and computational statistics to improve health care through predictive models and individualized treatment. Jordan Boyd-Graber (University of Maryland) discussed qualities that ubiquitous machine learning should have to allow for a future filled with “natural” interactions with humans. He explained the use of question-answering artificial intelligence (AI) as a way of evaluating how well AI systems can communicate what they are “thinking” to humans.