Data modification can introduce artificial information. It is often assumed that the resulting artefacts are detrimental to training, whilst being negligible when analysing models. We investigate these assumptions and conclude that in some cases they are unfounded and lead to incorrect results. Specifically, we show current shape bias identification methods and occlusion robustness measures are biased and propose a fairer alternative for the latter. Subsequently, through a series of experiments we seek to correct and strengthen the community's perception of how distorting data affects learning. Based on our empirical results we argue that the impact of the artefacts must be understood and exploited rather than eliminated.
MotivationModifying data has become commonplace both when training and analysing models, yet the wider implications are often disregarded. As examples of data modification, on the analysis side we take occlusion robustness and shape bias identification methods. On the training side, we focus on some instances of Mixed Sample Data Augmentation (MSDA), where two images are combined to obtain a new training sample. Visual illustrations of each can be found in Figure 1. In this paper we delve into some of the side-effects of data modification and point out that this practice has resulted in the creation of biased model interpretation tools and poorly informed theories. More specifically, we study a number of assumptions which we show are erroneous and which lie at the heart of the methods we briefly introduce below. Contesting these assumptions has broader implications on the community's perception of what aspects of the data are important when learning.Shape-texture bias: Deep models are known to be sensitive to interventions that are imperceptible to humans [35,14], as well as to other forms of distribution shifts [1,7,9]. It has been argued that this is intimately linked to networks tending to use texture rather than shape information [2,12]. Recently, input distortions have become a popular way of assessing a model's texture bias. To this end, images are divided into a grid and the resulting patches are randomly shuffled such that information is preserved locally, while the global shape is altered [32,26,24,42]. It is implicitly assumed that patch-shuffling does not introduce misleading shape or texture that could affect model evaluation.That is, if a model's accuracy drops when evaluated on patch-shuffled images, this degradation in performance is entirely attributed to the model's bias for shape information. Thus, any side-effects of the data manipulation process are considered negligible.Occlusion robustness: Commonly, occlusion robustness is concerned with the amount of information that can be hidden from a model without affecting its ability to classify [e.g. 36, 28]. A widely adopted proxy for measuring occlusion robustness is through the raw accuracy obtained after superimposing a rectangular patch on an image [6,10,40,43,21]. We refer to this approach as CutOcclusion throughout the paper...