@sknudstrup presents exciting new results showing that visual saliency exploration provides clear, interpretable ways to guide Monty’s sensors. This allows for learning models with greater sparsity without sacrificing accuracy. It also makes inference more efficient, allowing for recognition with fewer movements. This new policy has some nice synergies with the burst sampling that was recently added to Monty by @rmounir.
The model-free saliency policy can work in tandem with model-based policies, with the goal of having the sensory modules quickly gather relevant information to bring the learning modules in as soon as possible.
SalienceSM documentation: SalienceSM
Summary Video
Main Video
0:00 Introduction
2:04 Goal: Efficient Exploration
5:07 Goal: Sparser Models
7:45 Approach: Visual Salience
12:07 Approach: VOCUS 2
20:12 Results in Learning Models with Extra Sparsity
23:56 Inference Ability
33:49 Final Thoughts