GuessWhat?! A Visually grounded goal-oriented dialogue tasks for Deep Reinforcem

GuessWhat?! A Visually grounded goal-oriented dialogue tasks for Deep Reinforcement Learning

Résumé :

We introduce GuessWhat?!, a two-player guessing game as a testbed for research on the interplay of computer vision and dialogue systems. The goal of the game is to locate an unknown object in a rich image scene by asking a sequence of questions. Higher-level image understanding, like spatial reasoning and language grounding, is required to solve the proposed task.The goal of the game is to locate an unknown object in a rich image scene by asking a sequence of questions.

Higher-level image understanding, like spatial reasoning and language grounding, is required to solve the proposed task.Higher-level image understanding, like spatial reasoning and language grounding, is required to solve the proposed task.We will first present the dataset consisting of 150K human-played games with a total of 800K visual question-answer pairs on 66K images.AfterAfter introducing basic Deep Learning architecture, we will outline that such tasks to cannot only be trained in a supervised learning fashion.Détails :

One must cast the problem into the Reinforcement Learning framework to successfully render the intrinsic planning problem inherent to dialogues.

Bio :

Florian studies Deep Reinforcement Learning methods for Visual Dialogue Systems. After graduating from Ecole Des Mines de Saint-Etienne and Imperial College London in Machine Learning, he worked for two years in industry. He then became a research engineer at Inria for a year and started his Phd in the Inria SequeL team. He is also assistant professor in computer science at the University of Lille 1.

Lecture :

• https://arxiv.org/abs/1611.08481 - CVPR 2017

• https://arxiv.org/abs/1703.05423 - IJCAI 2017

Détails :

La présentation sera en français et est sponsorisé par Cognitalk.