ABSTRACT: Humans possess an impressive ability to process the relations between objects in a manner that abstracts away from irrelevant details about those objects. This is epitomized by the ability to perform analogies, mapping a relation between one pair of objects on to a superficially dissimilar pair of objects. Recently, a number of neural network models have been proposed to account for this ability. I will present results from a series of experiments evaluating one such model, the relation network. These experiments show that, although the model demonstrates state-of-the-art performance on challenging relational reasoning benchmarks, it falls short of the kind of strong generalization exhibited by human relational reasoning, highlighting the inherent difficulty of understanding relations at the right level of abstraction. I will conclude by considering possible future directions for evaluating and improving the generalization capabilities of relational reasoning models.