Sep 30, 2020 – 11:00 AM
In this seminar, some recent imitation learning algorithms will be presented. Imitation learning, also known as learning from demonstrations, is a powerful and practical alternative to reinforcement learning for learning sequential decision-making policies without the need of defining any (hand-crafted) reward function. We will present the GAIL algorithm, a model-free imitation learning algorithm based on a generative adversarial training.
In the second part of the seminar, following a recent paper by Z. Yang et al., this framework will be applied to the prediction of goal-directed scanpaths, with the aim of simulating human gaze behavior when searching for a specific target in a visual scene. The presented computational scheme outperforms other baseline models, both in terms of search efficiency and similarity to human search behavior. The learned reward maps reveal distinctive target-dependent patterns of object prioritization. Finally, spatial relationships between different objects are also well understood by the presented model.
– Yang, Zhibo, et al. “Predicting Goal-directed Human Attention Using Inverse Reinforcement Learning.” Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 2020.
– Ho, Jonathan, and Stefano Ermon. “Generative adversarial imitation learning.” Advances in neural information processing systems. 2016.