Interpretable CNNs for Object Classification

Abstract

This paper proposes a generic method to learn interpretable convolutional filters in a deep convolutional neural network (CNN) for object classification, where each interpretable filter encodes features of a specific object part. Our method does not require additional annotations of object parts or textures for supervision. Instead, we use the same training data as traditional CNNs. Our method automatically assigns each interpretable filter in a high conv-layer with an object part of a certain category during the learning process. Such explicit knowledge representations in conv-layers of CNN help people clarify the logic encoded in the CNN, i.e. answering what patterns the CNN extracts from an input image and uses for prediction. We have tested our method using different benchmark CNNs with various structures to demonstrate the broad applicability of our method. Experiments have shown that our interpretable filters are much more semantically meaningful than traditional filters.

CX-ToM: Counterfactual explanations with theory-of-mind for enhancing human trust in image recognition models

“…(B) Whereas in the second scenario (B), CX-ToM thinks that users do not trust the model's ability in correctly classifying Person, and therefore shows a faultline explanation using categories Woman and Deer. Biran and Cotton, 2017;Darlington, 2013;Kim, 2017a, 2017b;Goodman and Flaxman, 2017;Hoffman, 2017;Keil, 2006;Kulesza et al, 2010Kulesza et al, , 2011Moore and Swartout, 1990;Walton, 2004;Douglas, 2007;Walton, 2011;Sheh, 2017;Sheh and Monteath, 2018;Tapaswi et al, 2016;Williams et al, 2016;Agarwal et al, 2018;Akula et al, 2018Akula et al, , 2019aAkula et al, , 2019bAkula et al, , 2019cAkula et al, , 2021dAkula and Zhu, 2019;Gupta et al, 2016;Bivens et al, 2017;Zhang et al, 2019aZhang et al, , 2020aZhang et al, , 2020b. Most prior work in explaining CNN's predictions has focused on generating explanations using feature visualization and attribution.…”

COIN: Counterfactual Image Generation for Visual Question Answering Interpretation

“…Throughout the past decades, the notion of interpretability increasingly gained attention by the Machine Learning (ML) community [ 10 , 11 , 12 , 13 , 14 , 15 , 16 , 17 ]. According to Kim et al [ 13 ], interpretability is particularly important for systems whose decisions have a significant impact such as in healthcare, criminal justice and finance.…”

Explainability Metrics of Deep Convolutional Networks for Photoplethysmography Quality Assessment

“…We now direct our attention to local explanatory methods. One relevant class of locally explainable models is convolutional networks augmented with attention [ 27 ]. In such a case, the convolutional networks are trained to limit their vision to specific parts of the image, using network architecture which purposefully removes irrelevant information.…”