Perceptual Score: What Data Modalities Does Your Model Perceive?

Abstract

Machine learning advances in the last decade have relied significantly on largescale datasets that continue to grow in size. Increasingly, those datasets also contain different data modalities. However, large multi-modal datasets are hard to annotate, and annotations may contain biases that we are often unaware of. Deep-net-based classifiers, in turn, are prone to exploit those biases and to find shortcuts. To study and quantify this concern, we introduce the perceptual score, a metric that assesses the degree to which a model relies on the different subsets of the input features, i.e., modalities. Using the perceptual score, we find a surprisingly consistent trend across four popular datasets: recent, more accurate state-of-the-art multi-modal models for visual question-answering or visual dialog tend to perceive the visual data less than their predecessors. This trend is concerning as answers are hence increasingly inferred from textual cues only. Using the perceptual score also helps to analyze model biases by decomposing the score into data subset contributions. We hope to spur a discussion on the perceptiveness of multi-modal models and also hope to encourage the community working on multi-modal classifiers to start quantifying perceptiveness via the proposed perceptual score.Reported improvements are to a large extent due to the availability of large datasets [1-3], computational performance advances, e.g., for GPUs, and a better understanding about how to encode inductive biases into deep-nets, e.g., by using rectified linear units [4], normalization [5], skip connections [6], transformers [7], etc. However, importantly, developed deep-net architectures are not guaranteed to solve a given task. There is a chance that they may instead exploit dataset biases. This concern is surely in part due to non-robust training techniques, and a plethora of methods improve classifier robustness [8][9][10]. However, datasets play an important role in controlling the extracted bias as well. For instance, if correct answers in a question-answering task are significantly shorter than incorrect ones, classifier training should not use answer length as a cue. Although this seems reasonable, for audio-visual scene aware dialog, Schwartz et al. [11] find for example that in many cases the question alone is sufficient to generate a scene-aware dialog response, avoiding the need to look at the video. Hence, in order to assess the suitability of a classifier, we need to understand how much it relies on different data modalities.To quantify how much a classifier relies on its different input modalities, we introduce the perceptual score. The perceptual score assesses the degree to which a model relies on a modality. To do so the perceptual score permutes the features of a modality across samples in the test set after the classifier 35th Conference on Neural Information Processing Systems (NeurIPS 2021),