Malhar J

AI Research Assistant

Imperial College London • Full time

Jul 2016 - Sep 2016 • 2 mos

Worked as Research Assistant at Imperial College London. Currently Neural networks make incorrect predictions with high confidence (eg: high softmax output). This can be a big problem, especially in the context of safety-critical applications like autonomous vehicles where "trust" on the output is important. This project explored a method to prevent the network from making a prediction whenever it is uncertain (epistemic) about an input. Explored Bayesian methods for Predictive Uncertainty Estimation in Neural Networks in presence of ambiguous inputs. The task was to find a way to make the Neural network say “I know” or “I don’t know” on seeing an ambiguous input. For inputs that are classified as "I don't know", the network can refrain from making a prediction (and hence reduce confident but false predictions). Work involved Literature survey and prototyping a solution (using Tensorflow, Numpy, etc.). The literature survey involved reading up work on Bayesian Networks (MCDropout, GPDNN, BNN, CNN) and papers from important ML conferences like NIPS, ICML, etc.