B.S.H., M.S. Computer Science 2017 at Stanford University, specializing in Artificial Intelligence and Theory. My interests include making AI accessible to all potential application fields, increasing the interpretability of machine learning models, and increasing the robustness of AI models. My subject interests revolve around how theory from statistics, optimization, algorithms, and other math can be applied to machine learning. In my undergraduate career, I conducted research under Prof. Stefano Ermon in the Stanford Artificial Intelligence Laboratory, working on learning and inference for spatiotemporal machine learning models for computational sustainability and especially poverty reduction. I also helped to start the Sustainability and AI (SUSTAIN) lab at Stanford. At Lecida Inc., I work on time series prediction problems for sensor data from large industrial machines such as wind turbines, aiming to predict failures before they happen to reduce downtime.
Publicly available satellite imagery is used to accurately estimate local-level economic outcomes in five African countries.
We propose a novel machine learning approach to extract large-scale socioeconomic indicators from high-resolution satellite imagery. We utilize a fully convolutional neural network with transfer learning from a data-rich proxy task to learn features from satellite images of Africa to predict village-level poverty, approaching the performance of features from survey data.
Honors Thesis for the Bachelor's of Science with Honors Program at Stanford, detailing the transfer learning method for poverty mapping from satellite imagery, extending to multiple resolutions of satellite imagery, leveraging Gaussian Processes and deep kernel learning for semi-supervised learning to make use of abundant unlabeled data, and a general deployment pipeline for general data sources and spatial models, producing spatial maps. This scales the approaches in the previous papers, allowing us to provide global-scale poverty maps as satellite images are updated.
Deep learning techniques have led to massive improvements in recent years, but large amounts of labeled data are typically required to learn these complex models. We present a semi-supervised approach for training deep models that combines the feature learning capabilities of neural networks with the probabilistic modeling of Gaussian processes and demonstrate that unlabeled data can significantly improve performance on real-world datasets.
Presentation accompanying the AAAI-16 paper "Transfer Learning from Deep Features for Remote Sensing and Poverty Mapping".
Leveraging multiple resolutions of satellite imagery, we can cheaply take advantage of both the detail of high-resolution images and understand the context of the situation from low-resolution satellite imagery, which cover more land area. Both are important for predicting poverty measures, and we see that our multiple-resolution method enhances the state-of-the-art significantly across 5 African countries.
Fall 2015 - Use of a supervised pre-training step leveraging expert trajectories to augment the learning of a reinforcement learning agent through self-play. This idea was shown to be very successful in the Nature 2016 paper describing AlphaGo.
Generalization of the convolutional neural network encoder for the Encoder-Decoder paradigm in Neural Machine Translation proposed by Cho et. al. Cited by Facebook AI papers.
Deep learning applied to classifying images of plankton as a part of the National Data Science Bowl competition. As a part of the project, I developed 3 ways to sparsify probability distributions outputted by neural networks, using sampling techniques and convex optimization formulations.
Motion gesture recognition using Machine Learning on smart devices and smart watches.
We pioneered a new algorithm for dynamically truncating an accelerometer data stream for the important parts of the gesture.
Our classifier was able to achieve 96% accuracy on only 1 training example per gesture, which is better than HMM approaches and Dynamic Time Warping approaches seen in our literature review*! This has exciting applications for
new intuitive human input methods and automatic bio-tracking for life behaviors and patterns.
*uWave from Rice University, which used a DTW based approach, achieved 98% accuracy on 1 training example, using simple gestures such as moving left, moving right, circle, etc. Our gestures are arguably more complicated, which involves drawing the letters O, W, X, Z, and V in the air.
We built an extraction-based machine text summarizer that takes a piece of text and identifies the important sentences. This is a very challenging NLP problem; a machine learning approach was used to learn "importance" from a corpus of ~2000 documents.
My first Android app. Android application for background music listening on YouTube. The app has over 10000 downloads over its lifetime. YouTube has since taken the app down for providing a service which they provide through YouTube Red.
Rachmaninoff Piano Concerto No.1, 3rd Mvt
Chopin Sonata No.2, Mvt. 1
Chopin Sonata No.2, Mvt. 2
Rachmaninoff Piano Concerto No.1, 1st Mvt