Becks Simpson

Joseph D. Viviano, Becks Simpson, Francis Dutil et al.

Poor generalization is one symptom of models that learn to predict target variables using spuriously-correlated image features present only in the training distribution instead of the true image features that denote a class. It is often thought that this can be diagnosed visually using attribution (aka saliency) maps. We study if this assumption is correct. In some prediction tasks, such as for medical images, one may have some images with masks drawn by a human expert, indicating a region of the image containing relevant information to make the prediction. We study multiple methods that take advantage of such auxiliary labels, by training networks to ignore distracting features which may be found outside of the region of interest. This mask information is only used during training and has an impact on generalization accuracy depending on the severity of the shift between the training and test distributions. Surprisingly, while these methods improve generalization performance in the presence of a covariate shift, there is no strong correspondence between the correction of attribution towards the features a human expert has labelled as important and generalization performance. These results suggest that the root cause of poor generalization may not always be spatially defined, and raise questions about the utility of masks as "attribution priors" as well as saliency maps for explainable predictions.

Kyle Young, Gareth Booth, Becks Simpson et al.

Deep learning techniques have proven high accuracy for identifying melanoma in digitised dermoscopic images. A strength is that these methods are not constrained by features that are pre-defined by human semantics. A down-side is that it is difficult to understand the rationale of the model predictions and to identify potential failure modes. This is a major barrier to adoption of deep learning in clinical practice. In this paper we ask if two existing local interpretability methods, Grad-CAM and Kernel SHAP, can shed light on convolutional neural networks trained in the context of melanoma detection. Our contributions are (i) we first explore the domain space via a reproducible, end-to-end learning framework that creates a suite of 30 models, all trained on a publicly available data set (HAM10000), (ii) we next explore the reliability of GradCAM and Kernel SHAP in this context via some basic sanity check experiments (iii) finally, we investigate a random selection of models from our suite using GradCAM and Kernel SHAP. We show that despite high accuracy, the models will occasionally assign importance to features that are not relevant to the diagnostic task. We also show that models of similar accuracy will produce different explanations as measured by these methods. This work represents first steps in bridging the gap between model accuracy and interpretability in the domain of skin cancer classification.

Becks Simpson, Francis Dutil, Yoshua Bengio et al.

With too few samples or too many model parameters, overfitting can inhibit the ability to generalise predictions to new data. Within medical imaging, this can occur when features are incorrectly assigned importance such as distinct hospital specific artifacts, leading to poor performance on a new dataset from a different institution without those features, which is undesirable. Most regularization methods do not explicitly penalize the incorrect association of these features to the target class and hence fail to address this issue. We propose a regularization method, GradMask, which penalizes saliency maps inferred from the classifier gradients when they are not consistent with the lesion segmentation. This prevents non-tumor related features to contribute to the classification of unhealthy samples. We demonstrate that this method can improve test accuracy between 1-3% compared to the baseline without GradMask, showing that it has an impact on reducing overfitting.