1.1 Motivation 3 the true reasons why a datapoint has a certain label. In particular, such shortcuts may focus on unintended bias in the datasets. In a typical example by Ribeiro et al. (2016), a model is trained to differentiate huskies from wolves, but the dataset mainly contained wolves on a snowy background and huskies on other backgrounds. The model essentially learned a snow detector instead, predicting wolf whenever the background contained snow or a lot of white, thus incorrectly classifying images of huskies in the snow as wolves. There is often a lot of information in the data that isn’t described by a simple label fromY. E.g. animals may be recognised by the shape of their ears or the patterns in their fur, but the background of an image does not influence which animal is shown. Humans can identify such properties and use them to decide what animal they see, and they can articulate why they make this decision. Discriminative models mostly lack such explainability, since the task of predicting Y givenXdoesn’t encourage or require such information explicitly. High-dimensional data are typically observations from the real world, which is governed by certain underlying real-world mechanisms. To formalise this, we may say that world descriptors in a space of world states Wproduce observations in the data space X, see Figure 1.2. For example, a picture of an animal is an observation of that particular animal, in some pose, on some background. The animal itself can be described by properties such as the shape of its ears and nose, the colour and patterns of its fur or skin, and whether it has legs or wings. Such world descriptors in Ware not observed directly and can still be complicated, but they are typically much more low-dimensional and meaningful than data observations inX. Figure 1.2: Data inXare observations governed by underlying world descriptors inW. Discriminative modelling, i.e. directly predicting labels in Y from data in X, essentially sidesteps such world descriptors in W. However, from a good world description in W, the correct label in Y typically follows trivially and with a clear explanation, as illustrated schematically in Figure 1.3. Although world descriptors in W are generally unknown or unobserved, we are often able to model or extract certain aspects of it. Sidestepping Winherently limits
RkJQdWJsaXNoZXIy MjY0ODMw