Online Learning with Randomized Neural Networks

Arielle Anderer, PhD Candidate, The Wharton School; Hamsa Bastani, Operations, Information and Decisions, The Wharton School; Divya Singhvi, NYU Stern

Abstract: This project adapts online learning techniques to unstructured data, focusing on neural networks. We posit a non-parametric multi-armed bandit algorithm with context that uses a neural network for its underlying function approximation. We theoretically analyze and prove guarantees on the regret under this algorithm. Lastly, we examine its performance on a dataset of images of benign and malignant skin lesions, and compare it to that of other existing algorithms.