Further steps

• Implement skeleton based on stable-baselines (see https://github.com/Pastafarianist/rl-attention/issues/1)
• Test training of vanilla architecture locally and ensure that its performance is equivalent to published results on 2-3 environments
• Implement attention layer with architecture from LTIAA
• Implement visualisation of attention map & test it
• Test visualisation & performance of agent with attention
• Implement entropy loss
  • Apply to downsampled attention map
  • Apply to blurred attention map
• Add entropy loss to model with attention (experiment with different coefficients)
• Quantify difference in attention map between original model and model with entropy loss via Earth Mover Distance
• Implement gaussian mixture likelihood loss (make number of gaussians configurable)
• Add GM likelihood loss to model with attention (experiment with different coefficients)
• Quantify difference in attention map between original model and model with GM likelihood loss via Earth Mover Distance
• Implement Hooker et al.'s method for testing saliency and test attention model:
  • without modifications
  • with entropy loss
  • with GM likelihood loss