NeurIPS 2021

Timezone: »

Harmonizing Attention: Attention Map Consistency For Unsupervised Fine-Tuning

Ali Mirzazadeh · Florian Dubost · Daniel Fu · Khaled Saab · Christopher Lee-Messer · Daniel Rubin

in Bridging the Gap: from Machine Learning Research to Clinical Practice »

Learning meaningful representations is challenging when the training data is scarce. Attention maps can be used to verify that a model learned the target representations. Those representations should match human understanding, be generalizable to unseen data, and not focus on potential bias in the dataset. Attention maps are designed to highlight regions of the model’s input that were discriminative for its predictions. However, different attention maps computation methods often highlight different regions of the input, with sometimes contradictory explanations for a prediction. This effect is exacerbated when the training set is small. This indicates that either the model learned incorrect representations or that the attention maps methods did not accurately estimate the model’s representations. We propose an unsupervised fine-tuning method that optimizes the consistency of attention maps and show that it improves both classification performance and the quality of attention maps. We propose an implementation for two state-of-the-art attention computation methods, Grad-CAM and Guided Backpropagation, which relies on an input masking technique. We evaluate this method on our own dataset of event detection in continuous video recordings of hospital patients aggregated and curated for this work. As a sanity check, we also evaluate the proposed method on PASCAL VOC. On the video data, we show that the method can be combined with SimCLR, a state-of-the-art self-supervised training method, to further improve classification performance. With the proposed method, we achieve a 6.6 points lift of F1 score over SimCLR alone for classification on our video dataset, a 2.9 point lift of F1 score over ResNet for classification on

Author Information

Ali Mirzazadeh (Georgia Tech)

Florian Dubost (Stanford University)

I am a postdoctoral research fellow in Prof. Daniel Rubin's lab at Stanford. I develop neural networks to predict seizures from EEG and video recordings of epileptic patients and supervise students on related topics. During my PhD, I supervised over 15 international students, managed student teams and led them to top 3 positions in international deep learning competitions. My fields of expertise are engineering, computer science, AI, deep learning, machine learning, weakly supervised learning, self-supervised learning, image generation, registration, and variational autoencoders, with application in dementia, stroke, scoliosis, emphysema, cystic fibrosis, accelerated MRI reconstruction, and brain lesion detection. Articles include: Dubost, F., Adams, H., Yilmaz, P., Bortsova, G., van Tulder, G., Ikram, M.A., Niessen, W., Vernooij, M.W. and de Bruijne, M., 2020. Weakly supervised object detection with 2D and 3D regression neural networks. Medical Image Analysis, 65, p.101767. Dubost, F., Bortsova, G., Adams, H., Ikram, A., Niessen, W.J., Vernooij, M. and De Bruijne, M., 2017, September. GP-Unet: Lesion detection from weak labels with a 3D regression network. In International Conference on Medical Image Computing and Computer-Assisted Intervention (pp. 214-221). Springer, Cham. Dubost, F., Yilmaz, P., Adams, H., Bortsova, G., Ikram, M.A., Niessen, W., Vernooij, M. and de Bruijne, M., 2019. Enlarged perivascular spaces in brain MRI: Automated quantification in four regions. NeuroImage, 185, pp.534-544. More articles here: https://scholar.google.com/citations?user=_yNBmx8AAAAJ&hl=fr

Daniel Fu (Stanford University)

Khaled Saab (Stanford University)

Christopher Lee-Messer (Stanford University)

Daniel Rubin (Stanford University)

Dr. Rubin is a tenured Associate Professor of Biomedical Data Science, Radiology, and Medicine (Biomedical Informatics Research) at Stanford University. His NIH-funded research program focuses on artificial intelligence in medicine and quantitative imaging, integrating imaging with clinical and molecular data, and mining these Big Data to discover imaging phenotypes that can predict disease biology, define disease subtypes, and personalize treatment. Key contributions include discovering quantitative imaging phenotypes in radiology, pathology, and ophthalmology images that identify novel clinical subtypes of disease that help to determine treatments and improve clinical outcomes. He has over 240 peer-reviewed publications and 10 inventions.