PI: Asim Smailagic
Co-PI(s): Dan Siewiorek
University: Carnegie Mellon University
Explainable detection research in machine learning is a new and important area. We propose our method for the task of detecting diabetic retinopathy (DR) lesions from a dataset of eye fundus images containing a single scalar label for the whole image. DR is a worldwide leading cause of preventable blindness affecting more than 25% of the estimated 425 million diabetic patients in the world. Trained via weak supervision, our model pinpoints regions of the image containing lesions indicative of DR, thereby providing a highly valuable explanation when the model predicts presence of disease. Our method can convert any pre-trained convolutional neural network into a weakly-supervised model leading toward results showing the converted model provides both increased performance and efficiency. Next, we introduce a novel online active deep learning method for image analysis with a sampling method that queries the unlabeled examples maximizing the average distance to all training set examples. We will also experiment how domain knowledge that is encoded into a deep neural network by pre-training can improve the performance or and ability to explain a classification task. We will be producing attention maps at intermediate steps during the training process as to improve the ability to explain a task.