PREDICTION OF POTENTIAL HUMAN INTENTION USING SUPERVISED COMPETITIVE LEARNING

Masayoshi Ishikawa, Mariko Okude, Takehisa Nishida & Kazuo Muto

Nov 01, 2016 (modified: Nov 01, 2016) ICLR 2017 conference submission readers: everyone
  • Abstract: We propose a learning method to quantify human intention. Generally, a human being will imagine several potential actions for a given scene, but only one of these actions will subsequently be taken. This makes it difficult to quantify human intentions. To solve this problem, we apply competitive learning to human behavior prediction as supervised learning. In our approach, competitive learning generates several outputs that are then associated with several potential situations imagined by a human. We applied the proposed method to human driving behavior and extracted three potential driving patterns. Results showed a squared error is reduced to 1/25 that of a conventional method . We also found that competitive learning can distinguish valid data from disturbance data in order to train a model.
  • Conflicts: hitachi.com
  • Keywords: Computer vision, Deep learning, Supervised Learning, Applications

Loading