Minds vs. Machines: How far are we from the common sense of a toddler?

CVPR 2020 Workshop, June 15, Seattle, WA (room to be announced)


Invited Speakers

Alison Gopnik (Berkeley) Jitendra Malik (Berkeley) Aude Oliva (MIT)
Elizabeth Spelke (Harvard) Joshua Tenenbaum (MIT) Daniel Yurovsky (CMU)
Larry Zitnick (FAIR)

Schedule


Introduction

What can a toddler do? Although young toddlers might seem helpless, they have a basic understanding of how the world works (i.e., intuitive physics), how people work (i.e., intuitive psychology), and of what their parents tell them. Furthermore, they have learned these abilities without 3D bounding box or segmentation annotations, or annotations regarding goals and intentions. What they can do is so elementary that we often take it for granted. Yet, it remains elusive for current machine learning models for perception, language understanding, reasoning or interaction with the world.

Current AI systems do well in detecting and naming objects in photographs, recognizing actions in sports from YouTube videos, or answering complicated questions regarding images---questions they have been trained to answer. However, they cannot easily extrapolate their knowledge to new situations, they cannot reason about space and object locations, or about goals and intentions the way toddlers do. In short, they do not have common sense. Without common sense, our systems are unpredictable in unseen situations, are difficult to teach and communicate with, and do not self-improve in a stable manner.

In this workshop, we will try to answer the following questions:

  1. How far are current AI systems from the vision, language and reasoning abilities of a toddler?
  2. What are some insights we can draw from our understanding of toddlers and the human brain to improve current AI systems?
  3. To build human-like common sense, what research topics need continued exploration, and what topics are still missing?

We would like to bring together leading researchers on neuroscience, psychology, computer vision and robotics to discuss these questions and debate on their answers. We have an exciting list of invited speakers from these domains. We will also invite researchers to submit peer-reviewed papers on the aforementioned topics. Our one-day workshop will have a poster session, an oral session and a panel discussion to enable dialogue and the exchange of ideas.


Organizers

Katerina Fragkiadaki (CMU) Adam Harley (CMU) Phillip Isola (MIT) Fuxin Li (Oregon State)
Fish Tung (CMU) Aria Wang (CMU) Leila Wehbe (CMU) Jiajun Wu (Stanford)