CNC seminár z umelej inteligencie - Igor Farkaš (12.3.2025)

v stredu 12.3.2025 o 11:30 hod. v miestnosti I/8 aj online formou

---

Prednášajúci: prof. Ing. Igor Farkaš, Dr.

Názov prednášky: Self-supervised network distillation for efficient learning sequential tasks with sparse reward

Termín: 12.3.2025, 11:30 hod., I 8 aj MS Teams

Abstrakt: 
Reinforcement learning can solve decision-making problems by training an agent to behave in an environment according to a predesigned reward function. This is, however, very difficult if the reward is too sparse, so the solution may be to equip the agent with an intrinsic motivation (IM) that will provide informed exploration during learning. Novelty detection is one of the promising branches in this domain. We present Self-supervised Network Distillation (SND), a class of IM algorithms based on the distillation error as a novelty indicator, where the predictor model and the target model are both trained. We adapted three existing self-supervised methods and experimentally tested them on a set of ten environments that are considered difficult to explore. The results show that our approach is more effective compared to the baseline models. We also applied the analytical methods that help explain the virtues of our models. The presentation refers to our recent work https://arxiv.org/abs/2302.11563

---