Chess Game Concepts Emerge under Weak Supervision: A Case Study of Tic-tac-toe

Hao Zhao, Ming Lu, Anbang Yao, Yurong Chen, Li Zhang

Nov 04, 2016 (modified: Dec 13, 2016) ICLR 2017 conference submission readers: everyone
  • Abstract: This paper explores the possibility of learning chess game concepts under weak supervision with convolutional neural networks, which is a topic that has not been visited to the best of our knowledge. We put this task in three different backgrounds: (1) deep reinforcement learning has shown an amazing capability to learn a mapping from visual inputs to most rewarding actions, without knowing the concepts of a video game. But how could we confirm that the network understands these concepts or it just does not? (2) cross-modal supervision for visual representation learning draws much attention recently. Is this methodology still applicable when it comes to the domain of game concepts and actions? (3) class activation mapping is widely recognized as a visualization technique to help us understand what a network has learnt. Is it possible for it to activate at non-salient regions? With the simplest chess game tic-tac-toe, we report interesting results as answers to those three questions mentioned above. All codes, pre-processed datasets and pre-trained models will be released.
  • TL;DR: investigating whether a CNN understands concepts from a new perspective
  • Keywords: Semi-Supervised Learning
  • Conflicts:,