Building and Interpreting Deep Similarity Models

Abstract

Many learning algorithms such as kernel machines, nearest neighbors, clustering, or anomaly detection, are based on the concept of 'distance' or 'similarity'. Before similarities are used for training an actual machine learning model, we would like to verify that they are bound to meaningful patterns in the data. In this paper, we propose to make similarities interpretable by augmenting them with an explanation in terms of input features. We develop BiLRP, a scalable and theoretically founded method to systematically decompose similarity scores on pairs of input features. Our method can be expressed as a composition of LRP explanations, which were shown in previous works to scale to highly nonlinear functions. Through an extensive set of experiments, we demonstrate that BiLRP robustly explains complex similarity models, e.g. built on VGG-16 deep neural network features. Additionally, we apply our method to an open problem in digital humanities: detailed assessment of similarity between historical documents such as astronomical tables. Here again, BiLRP provides insight and brings verifiability into a highly engineered and problem-specific similarity model.

Towards Robust Explanations for Deep Neural Networks

“…An interesting direction for future research will be to relate the established limits of robust explanation methods to techniques for uncertainty quantification respectively in relation to methods studying the relevant structural parts in learning models [56,57]. Furthermore it will be helpful to discuss resilience to manipulation of explanation methods also in the context of unsupervised learning [58][59][60] and multi-modal data/similarity streams [61].…”

Improving a neural network model by explanation-guided training for glioma classification based on MRI data

“…the technique named BiLRP [8] to show common features between an input and sample atlas images. This explanation provides higher-order explanations that extend conventional visualization techniques.…”

“…Here we proposed to use higher-order explanations to show domain experts what features between an input image and an image from ground truth cases are commonly similar. To bring these higher-order explanations to MRI data classification, we introduce using of BiLRP [8] technique. Similar atlas images to the input image are visualized in our method.…”

“…We can detect adversarial examples or protect model from examples that are outside of model understanding. BiLRP[8]…”

Evolution and transformation of early modern cosmological knowledge: a network study

“…We will continue the analysis of the corpus by considering other "knowledge atoms," namely scientific images and computational astronomic tables extracted from the same textbooks using machine learning techniques [34].…”