Understanding Grounded Language Learning Agents

[kweonwooj]

Abstract

• Learning words by children is fascinating
• Investigate HOW agent learns the grounded language learning in neural-network based agent in simulation

Details

• Model Architecture and Environment
  • CNN to observe environment + LSTM for question + Merge + LSTM to generate answer

Four main lessons

• Shape/Color bias

  • when learning colors or shape only, bias is skewed to color or shape respectively (bias in color is stronger)
  • when equal number of color/shape is learned, bias is skewed to color! (agent favors learning color, because it is more simpler to find pattern)

• Learning Negation

  • generalizing negation is difficult, it requires more examples

• Curriculum Effect for Vocab Growth
  • curriculum learning makes sense in RL

• Representation Difference
  • learns different semantic classes at different speeds and represent them with features that require different degrees of visual processing depth

• Layer-wise Attention
  • each CNN layer attending to different aspects

Personal Thoughts

• Good bridge between machine learning model and cognitive psychology
• why is curriculum learning not effective in NMT?
• envy how DeepMind can do this exploratory analytical paper

Link : https://arxiv.org/pdf/1710.09867.pdf Authors : Hill et al. 2017