20 questions approach could cut AI training costs

University of Bristol-led research suggests image-classification systems could be trained more cheaply by combining many simple binary classifiers into a sequence of yes/no decisions inspired by 20 Questions. Presented at the Allerton Conference on Communication, Control, and Computing in Illinois, in the US on Wednesday 16 September, the method uses classifiers that can be trained in a matter of minutes on a standard laptop, rather than systems that normally require tens of thousands of Graphics Processing Units and cost millions of dollars.

Prof Sidharth Jaggi said the work mathematically proves, with empirical validation, that simple random questions can be combined to handle complex classification tasks. The approach does not require complex coordination between binary classifiers, only enough of them, and the number needed is described as surprisingly small.

The method is positioned for AI systems running across multiple devices or directly inside smart devices, including sensors, robots, and edge devices that process data close to where it is generated. Because each question is answered independently, the overall system can still produce reliable results even when some individual answers are wrong.

The work forms part of the Informed AI research hub at the University of Bristol, which focuses on mathematics, information theory, and AI safety. The hub aims to support UK AI systems that are efficient, robust, reliable, safe, trustworthy, and suitable for deployment in society.