Machine learning for a sustainable world
Abstract/Contents
- Abstract
- In 2015, the United Nations General Assembly set 17 Sustainable Development Goals to serve as the guiding principles for international development over the next 15 years. This thesis explores how machine learning could help to address some of these global sustainability challenges. The first part of the thesis focuses on data gaps in the developing world that make it hard to measure progress and target intervention effectively. Traditional data collection methods like household surveys are slow and expensive. Combining machine learning with passively collected remote sensing data could prove to be a scalable alternative, but a lack of labeled data poses a major challenge for sustainability applications. To combat this data scarcity, we propose a transfer learning approach that uses nighttime lights as a proxy for economic development. By extracting predictive features from daytime satellite imagery, we can generate fine-grained poverty and wealth estimates and create high-resolution maps of poverty. Next, we present semi-supervised deep kernel learning (SSDKL) to leverage the large amounts of unlabeled satellite data. We demonstrate that SSDKL learns more generalizable features and improves performance on a range of semi-supervised regression tasks. Finally, we introduce Tile2Vec, an unsupervised representation learning algorithm. We evaluate Tile2Vec on a wide range of remote sensing datasets, and show that it even works on non-image spatial data. The second part of the thesis explores culture-free diagnostics for bacterial infections, a leading cause of death in developing nations. Current diagnostic methods require sample culturing to identify the bacteria and its antibiotic susceptibility, a slow process that can take days even in state-of-the-art labs. Broad spectrum antibiotics are often prescribed while waiting for culture results, leading to suboptimal therapy and contributing to the increased prevalence of antibiotic resistance. We present a proof-of-concept system that combines Raman spectroscopy and deep learning to achieve accurate bacterial identification and susceptibility testing in a single step. We generate an extensive dataset of bacterial Raman spectra, and show that we can accurately identify 30 common bacterial pathogens. Even on low signal-to-noise spectra, we achieve antibiotic treatment identification accuracies of 97% and distinguish between methicillin-resistant and -susceptible isolates of Staphylococcus aureus (MRSA and MSSA) with 89% accuracy. Finally, we validate our results on clinical samples spanning 50 patients, where we achieve treatment identification accuracies of 99.7%.
Description
| Type of resource | text |
|---|---|
| Form | electronic resource; remote; computer; online resource |
| Extent | 1 online resource |
| Place | California |
| Place | [Stanford, California] |
| Publisher | [Stanford University] |
| Copyright date | 2019; ©2019 |
| Publication date | 2019; 2020 |
| Issuance | monographic |
| Language | English |
Creators/Contributors
| Author | Jean, Neal |
|---|---|
| Degree supervisor | Ermon, Stefano |
| Thesis advisor | Ermon, Stefano |
| Thesis advisor | Johari, Ramesh, 1976- |
| Thesis advisor | Sadigh, Dorsa |
| Degree committee member | Johari, Ramesh, 1976- |
| Degree committee member | Sadigh, Dorsa |
| Associated with | Stanford University, Department of Electrical Engineering. |
Subjects
| Genre | Theses |
|---|---|
| Genre | Text |
Bibliographic information
| Statement of responsibility | Neal Jean |
|---|---|
| Note | Submitted to the Department of Electrical Engineering |
| Thesis | Thesis Ph.D. Stanford University 2020 |
| Location | electronic resource |
Access conditions
- Copyright
- © 2019 by Neal Jean
- License
- This work is licensed under a Creative Commons Attribution Non Commercial 3.0 Unported license (CC BY-NC).
Also listed in
Loading usage metrics...