Recurrent Model of visual Attention

• Pytorch implementation of Mnih, Volodymyr, Nicolas Heess, and Alex Graves. "Recurrent models of visual attention." Advances in neural information processing systems. 2014.

Recurrent Attention Model

Implementation Details

• RAM for classification is defined in RAM.py
  • The core network is recurrent network
  • Action network is trained with supervised objective
  • Location network is trained with REINFORCE
• Individual experiments are implementated in RAM_centered.py,RAM_translated.py and RAM_cluttered.py
• Environment for game Catch is defined in catch.py
• RAM and experiment for Catch are defined in RAM_catch.py
  • The core network is LSTM
  • Action and location networks are trained with REINFORCE

Tasks

• RAM is tested in 4 scenarios:
  • Classification of centered MNIST digit (size: 28x28)
  • Classification of translated MNIST digit (size: 60x60)
  • Classification of cluttered and translated MNIST digit (size: 60x60, 4 pieces of clutter of size 8x8)
  • Playing game Catch

Classification of MNIST digits

Experiments

Hyper-parameters	Centered digit	Translated digit	Cluttered and Translated digit
Glimpse size	8x8	12x12	12x12
Glimpse scale	1	3	3
Initial learning rate	0.001	0.0001	0.0001
Learning rate decay	0.95	0.975	0.975
std of location policy	0.25	0.25	0.25
Epochs	100	200	200

Performance

Number of glimpses	Centered digit	Translated digit	Cluttered and Translated digit
2	89.54	-	-
3	94.48	-	-
4	96.82	89.35	83.00
5	97.09	-	-
6	97.54	91.74	85.50
7	97.57	-	-
8	-	92.61	85.46

Visualization

Visualization of policy on non-centered digit classification task for T=8. (a) Digit with gaze track (b) individual glimpses. Each column shows single glimpse with patches of decreasing scale from top to bottom. Intermediate predictions are shown on the top.

Playing Catch

Experiments

Hyper-parameters	values
Glimpse size	6x6
Glimpse scale	3
Initial learning rate	0.0001
Learning rate decay	0.99
std of location policy	0.25
Training episodes	15 million

Performance

Agent achieves ~89% accuracy.

Visualization

• Agent playing the game below

• Gaze-track on recency-based weighted average of frames from the game