Title: Visually Grounded Language Understanding and Generation

Jiasen Lu

Ph.D. Candidate in Computer Science

School of Interactive Computing

Georgia Institute of Technology

https://www.cc.gatech.edu/~jlu347/

Date: Monday, January 6, 2020

Time: 12:00-2:00 PM (EST)

Location: CODA C1108

BlueJeans: https://bluejeans.com/7313234985

Committee:

Dr. Devi Parikh (Advisor), School of Interactive Computing, Georgia Institute of Technology

Dr. Dhruv Batra, School of Interactive Computing, Georgia Institute of Technology

Dr. Mark Riedl, School of Interactive Computing, Georgia Institute of Technology

Dr. Judy Hoffman, School of Interactive Computing, Georgia Institute of Technology

Dr. Jason J. Corso, Department of Electrical Engineering and Computer Science University of Michigan 

Abstract:

The world around us involves multiple modalities. One of the major challenges in modeling different modalities jointly is how to induce appropriate grounding in models given the heterogeneity of the data. Which parts of the image and question should the model focus on when answering a question about an image? How can we integrate object detectors to produce fluent but visually grounded image captions? How can we disentangle "what to say" from "how to say it" when automatically generating goal-oriented dialogs about images? How to build a more general multi-modal AI that can learn visual groundings from massive meta-data on the internet and solve multiple tasks at the same time. 

In this thesis, I take steps towards studying how inducing appropriate grounding in deep models improves multi-modal AI capabilities, in the context of vision and language understanding.

Specifically, I will present --

1) how to ground visual question answering models in appropriate regions of the image and appropriate phrases in the question to more accurately answer questions about images

2) how to ground image captioning models in object detections by combining symbolic and deep learning approaches to avoid hallucinations of visual concepts in image captions

3) how to generalize from single round visual question generation with full supervision to a multi-round dialog-based image guessing game without direct language supervision.  

4) how to learn the joint visual-linguistic representations with self-supervised learning which have captured rich semantic and structural information from a large, unlabeled data source. 

On these vision-and-language tasks, I will demonstrate that inducing appropriate grounding in deep models improves multi-modal AI capabilities. To the end, I will briefly discuss the challenges in this domain and the extensions of recent works.