Teaching a neural network to use a calculator

This article explores a seq2seq architecture for solving simple probability problems in Saxton et. A transformer is used to map questions to intermediate steps, while an external symbolic calculator evaluates intermediate expressions. This approach emulates how a student might solve math problems, by setting up intermediate equations, using a calculator to solve them, and using those results to construct further equations. A few months ago, DeepMind released Mathematics Dataset, a codebase for procedurally generating pairs of mathematics questions and answers, to serve as a benchmark for the ability of modern neural architectures to learn mathematical reasoning. The data consists of a wide variety of categories, ranging from basic arithmetic to probability. Both questions and answers are in the form of free-form text, making seq2seq models a natural first step for solving this dataset.