Casimir J. H. Ludwig

Milton L. Montero, Jeffrey S. Bowers, Rui Ponte Costa et al.

Recent research has shown that generative models with highly disentangled representations fail to generalise to unseen combination of generative factor values. These findings contradict earlier research which showed improved performance in out-of-training distribution settings when compared to entangled representations. Additionally, it is not clear if the reported failures are due to (a) encoders failing to map novel combinations to the proper regions of the latent space or (b) novel combinations being mapped correctly but the decoder/downstream process is unable to render the correct output for the unseen combinations. We investigate these alternatives by testing several models on a range of datasets and training settings. We find that (i) when models fail, their encoders also fail to map unseen combinations to correct regions of the latent space and (ii) when models succeed, it is either because the test conditions do not exclude enough examples, or because excluded generative factors determine independent parts of the output image. Based on these results, we argue that to generalise properly, models not only need to capture factors of variation, but also understand how to invert the generative process that was used to generate the data.

Cindy Xiong, Cristina R. Ceja, Casimir J. H. Ludwig et al.

In visual depictions of data, position (i.e., the vertical height of a line or a bar) is believed to be the most precise way to encode information compared to other encodings (e.g., hue). Not only are other encodings less precise than position, but they can also be prone to systematic biases (e.g., color category boundaries can distort perceived differences between hues). By comparison, position's high level of precision may seem to protect it from such biases. In contrast, across three empirical studies, we show that while position may be a precise form of data encoding, it can also produce systematic biases in how values are visually encoded, at least for reports of average position across a short delay. In displays with a single line or a single set of bars, reports of average positions were significantly biased, such that line positions were underestimated and bar positions were overestimated. In displays with multiple data series (i.e., multiple lines and/or sets of bars), this systematic bias still persisted. We also observed an effect of "perceptual pull", where the average position estimate for each series was 'pulled' toward the other. These findings suggest that, although position may still be the most precise form of visual data encoding, it can also be systematically biased.