Machine Learning, 2nd Edition

Chapter 13Current Challenges in Machine Learning

Decision makers need forecasts only if there is uncertainty about the future.

— J. Scott Armstrong, Principles of Forecasting, 2001

This chapter briefly presents a few growth areas in machine learning, generally arising from the changing relative costs of acquiring data, transmitting data, storing data, analyzing data, and computing with data.

13.1 Streaming Data

Streaming data are data which are generated faster or in greater quantity than they can be put in long-term storage. Each datum, on arrival, must be used immediately for whatever purpose is appropriate (training a model, used for prediction, updating a clustering, etc.) and then discarded forever. One area of concern is the ongoing training and assessment of predictive models.

13.2 Distributed Data

Distributed data abstractly comprise a single dataset but physically reside in separate storage devices on a communications network of such high latency that it is infeasible for all of the data, in raw form, to be processed by a single processor. Data may be distributed because it was originally collected in multiple locations, or to reduce access time to a large amount of data by parallelizing the read operation across many hard drives. To be practical, algorithms applied to distributed data may only transmit across the network relatively small summaries of the data on any particular storage device. This is the basis of cloud computing.

13.3 Semi-Supervised Learning ...