Measuring thresholds is an established and time-proven method to obtain the minimal differences required to reliable discriminate stimuli: the just-noticeable difference (JND). However, in perception we are often times not only interested in the JND, but in subjective aspects of perception, for example, in how bright a light appears, or which textured surfaces appear more similar to one another. To obtain answers to such supra-threshold questions we typically employ scaling or magnitude-estimation techniques. With scaling techniques the perceived similarity relations between stimuli are mapped to a geometry, in which the distances in the internal (typically Euclidean) space correspond to the perceived similarities: Stimuli perceived to be similar should be close together in the putative internal representation, whilst stimuli perceived to be dissimilar should be far apart.

A number of scaling techniques exist, but in recent years ordinal embedding techniques from machine learning have been successfully used to infer internal representations from ordinal triplet comparisons in psychology. In my presentation I will explain the fundamentals of ordinal embedding techniques and argue that, together with ordinal triplet comparisons, ordinal embedding is a useful and reliable method to infer internal representations. For example, and unlike it is the case for other scaling techniques, for ordinal embeddings we have statistical means to objectively determine the dimensionality of the embedding. In addition, I will show that mathematically similar triplet question variants---"standard" triplets versus odd-one-out---are psychologically not the same.