Using Generative Models for Faces to Test Neural Networks (Talk)
Most machine learning models are validated on fixed datasets. This can give an incomplete picture of the capabilities and weaknesses of the model. Such weaknesses can be revealed at test time in the real world with dire consequences. In order to alleviate this issue, simulators can be controlled in a fine-grained manner using interpretable parameters to explore the semantic image manifold and discover such weaknesses before deploying a model. Also, in recent years there have been important advances in generative models for computer vision resulting in realistic face generation and manipulation. In this presentation I will show how face recognition neural networks can be tested using computer graphics-based simulators as well as contemporary realistic generative adversarial networks.
Biography: Nataniel is a PhD candidate at Boston University advised by Professor Stan Sclaroff. He obtained his Masters at Georgia Tech under James Rehg and his Bachelors at Ecole Polytechnique in Paris, France. He is currently a research intern at Amazon working with Javier Romero and has been a research intern at Apple AI Research, NEC-Labs and MIT. He has been selected as a 2020 Twitch Research Fellowship finalist and was a second round interviewee for the 2020 Open Phil AI Fellowship. His work has been featured on the TWIML AI podcast, Boston University’s The Brink newsletter, the Toward Data Science online publication, front page of Y-Combinator’s Hacker News and has been mentioned on Forbes magazine. His current research is around two main topics: (1) Methods to protect one's image and likeness by developing adversarial attacks to disrupt deepfake generation. (2) Methods to train and test machine learning models using simulated data in order to improve models and understand how they learn and generalize.
