Companies can sponsor the speaker series by sponsoring Duke Undergraduate Machine Learning. MLBytes speakers from previous semesters can be found here.
- Thursday, October 29, 2020: MLBytes Speaker Series Doubleheader: Applied Intuition + CMU
- 3:15 - 4:15: Synthetic Data Simulation for Autonomous Driving by Applied Intuition (Shrey Gupta, Duke '20 and Katherine Guo, Duke '19)
- Engineers and researchers working on autonomous driving are developing complex models to put self-driving cars on the road, but these models are hard to test. In particular, edge cases, by definition, are difficult to find, since they are rare. This talk describes how simulation and re-simulation using synthetic data are being used to test models and improve the generalization of various components of the autonomous driving stack, including planning and perception. Applied Intuition provides software infrastructure to safely develop, test, and deploy autonomous vehicles at scale, including using simulation and synthetic data.
- 4:30 - 5:30: Values and Tradeoffs in Learning from Consumer Location Data (Meganath Macha, PhD Candidate at Carnegie Mellon University)
- Location data has changed the way we understand human behavior. 76% of the population in the advanced economies own a smartphone. These percentages continue to rocket. The fast penetration of smartphones, combined with the wide adoption of location services, has produced a vast volume of behavior-rich mobile consumer location data. This talk presents three case studies in which we propose novel methods to discern the value and trade-offs in learning consumer behavior from location data.
First, we study the explainability and predictive accuracy trade-off in learning from location data. We present x-PACS, a new sub-space search learning algorithm that jointly explains and detects anomalous patterns. Explanations are useful in making learning algorithms more transparent to the data practitioner. Benchmarking on several real-world datasets, we show the effectiveness of x-PACS in anomaly explanation over various baselines and demonstrate its competitive predictive performance.
Second, we study the privacy-utility trade-off associated with the collection and sharing of consumer location data. We find that high privacy risks prevail in the absence of obfuscation on the shared location data. We propose a novel framework enabling a data collector to balance the privacy-utility trade-off. We empirically demonstrate the performance of this approach on smartphone location data of 40,000 consumers collected across several weeks.