I’m a 4th year PhD candidate in the Stanford Machine Learning Group co-advised by Professor Andrew Ng and Professor Percy Liang. I work on the development and deployment of deep learning algorithms for automated diagnosis, prognosis, and treatment of diseases. I have developed models for automated detection of arrhythmias, multiple pathology detection under uncertainty for x-rays (CheXNet, MURA, CheXNeXt), and augmentation of experts in knee MRI interpretation (MRNet). I have also developed SQuAD, a machine reading comprehension dataset.
We developed an algorithm to predict abnormalities in knee MRI exams, and measured the clinical utility of providing the algorithm’s predictions to radiologists and surgeons during interpretation. 7 practicing board-certified general radiologists and 2 practicing orthopedic surgeons at Stanford University Medical Center (3–29 years in practice, average 12 years) read a validation set of 120 exams twice, once without model assistance and once with model assistance, separated by a washout period of at least 10 days. We found that model assistance significantly reduced the rate at which healthy patients would be mistakenly diagnosed as having ACL tears.
We developed CheXNeXt, a deep learning algorithm to concurrently detect 14 clinically important diseases in chest radiographs. CheXNeXt's training process consists of 2 consecutive stages to account for the partially incorrect labels in the ChestX-ray14 dataset. We evaluated the algorithm against 9 practicing radiologists on a validation set of 420 images for which the majority vote of 3 cardiothoracic specialty radiologists served as ground truth. The algorithm achieved performance equivalent to the practicing radiologists on 10 pathologies, better on 1 pathology, and worse on 3 pathologies.
MURA is a large dataset of bone X-rays. Algorithms are tasked with determining whether an X-ray study is normal or abnormal. Musculoskeletal conditions affect more than 1.7 billion people worldwide, and are the most common cause of severe, long-term pain and disability, with 30 million emergency department visits annually and increasing. We hope that our dataset can lead to significant advances in medical imaging technologies which can diagnose at the level of experts, towards improving healthcare access in parts of the world where access to skilled radiologists is limited.
SQuAD2.0 combines the 100,000 questions in SQuAD1.1 with over 50,000 new, unanswerable questions written adversarially by crowdworkers to look similar to answerable ones. To do well on SQuAD2.0, systems must not only answer questions when possible, but also determine when no answer is supported by the paragraph and abstain from answering. SQuAD2.0 is a challenging natural language understanding task for existing models, and we release SQuAD2.0 to the community as the successor to SQuAD1.
We develop an algorithm that can detect pneumonia from chest X-rays at a level exceeding practicing radiologists. Our model, CheXNet, is a 121-layer convolutional neural network that inputs a chest X-ray image and outputs the probability of pneumonia along with a heatmap localizing the areas of the image most indicative of pneumonia. We train on ChestX-ray14, the largest publicly available chest X-ray dataset. We find that the model exceeds the average radiologist performance at the pneumonia detection task on both sensitivity and specificity.
Our deep learning algorithm exceeds the performance of board certified cardiologists in detecting a wide range of heart arrhythmias from electrocardiograms recorded with a single-lead wearable monitor. Dataset 500x larger than previously studied corpora used to train a deep convolutional neural network.
Stanford Question Answering Dataset (SQuAD) is a reading comprehension dataset, consisting of questions posed by crowdworkers on a set of Wikipedia articles, where the answer to every question is a segment of text, or span, from the corresponding reading passage. With 107,785 question-answer pairs on 536 articles, SQuAD is significantly larger than previous reading comprehension datasets.
A single class can transform a life. A popular introduction to programming class leads to the discovery of a passion for computer science, a social dance class exposes a deep appreciation for artistic expression--experiences like these are at the core of a Stanford education. Yet out of the 5000 classes offered here, students only have time to take less than 1% during their undergraduate career. This small selection of classes determines the foundation on which passions are developed - passions that lead to great innovations and great discoveries that change the world. Some students arrive at Stanford with clear visions of their futures. Others need a little time to explore and decide what to do with their lives. Edusalsa lets students find the classes where they can discover their passions, equipping them with new tools on their path of intellectual discovery, infusing life and vitality into the Stanford experience.
Research in Autonomous Driving spanning Computer Vision, Artificial Intelligence, and Crowdsourcing. My undergraduate honors research introduced Driverseat, a technology for embedding crowds around learning systems for autonomous driving.
Piloted for the Deep Learning Symposium at NIPS '15, Recommend-Papers was built in order to facilitate discussion of the most recent deep learning breakthroughs and explore an alternative mechanism for selecting presentations. In order to let the broader research community (including the authors of research papers) contribute to the discussion, Recommend-Papers allowed members to post papers and comment on them, and PC members to hold private discussions.
Can a computer identify the chord I'm playing on a guitar simply by listening to it? How well does machine learning perform on the task realtime? Could we leverage that technology to give realtime feedback to an instrument learner? This research presents a prototype of an online tool for real-time chord recognition. It fuses traditional techniques in machine learning with the capabilities of new web technologies such the the Web Audio API, and WebSockets.
Machine Learning Experiments (mlx) is a blog to showcase machine learning work intended to showcase machine learning experiments not just in their final polished form, but also highlight the thought process that guides research.
Research in Human Computer Interaction, and Natural Language Processing, exploring how we could teach a computer enough about human actions to enable predictive application interfaces that could, for example, recommend ice cream shops upon learning that a person was having dinner.
Singing is awesome, powerful, and personal. Can we simplify, for amateur singers, the process of exploring new songs to sing along to? Vocalet provides a simple interface for singing enthusiasts to enjoy. It's easy to sing along to karaoke versions of songs, and get inspired by cover artists.