Uncommon Sense: Using Neural Networks for Exploration and Creativity

Speaker Name: 
Genevieve Patterson
Speaker Title: 
Postdoctoral Researcher
Speaker Organization: 
Microsoft Research New England
Start Time: 
Wednesday, February 14, 2018 - 11:00am
End Time: 
Wednesday, February 14, 2018 - 12:15pm
Location: 
E2-506
Organizer: 
Lise Getoor

Abstract:

In this talk, Patterson investigates new ways for humans and AI systems to collaborate. She is interested in problems where pure automation is not the goal. For example, deep networks can automatically detect the presence of cancer in a mammogram but can’t tell us what avenues we should pursue in cancer research. She will discuss two recent projects where state-of-the-art AI fails users. Patterson will present methods for identifying the shortcomings of canonical AI solutions and changing optimized objectives to suit user needs. Her first project will demo a system for ML supported rotoscoping, an expensive post-processing technique commonly used in filmmaking. Her second project explores collaboration between DNNs and radiologists to generate explanations of neural network diagnoses. She’ll conclude with a quick introduction to her current work using deep learning to help people write better jokes. Through these projects, she hopes to demonstrate novel interactions between AI and humans.

Bio:

Genevieve Patterson studies interactive Computer Vision systems. She is a postdoctoral researcher at Microsoft Research New England. Her interests include visual phenomena discovery, crowd-driven dataset annotation, fine-grained object recognition, low-shot learning, and domain adaptation. Her work has appeared at several top tier computer vision venues including CVPR and ECCV, as well as HCI venues CHI and HCOMP, where her paper introducing a system for crowdsourced one-shot object detection was finalist for Best Paper in 2015. She received her PhD from Brown University in 2016 under the direction James Hays (now of Georgia Tech).