Meeting Details
Date : 23 March 2021
Time : 10:00 – 11:00
Location : Chair of CPS, Montanuniverität Leoben
Participants: Univ.-Prof. Dr. Elmar Rueckert, Fotios Lygerakis
Agenda
- Discuss idea for adding contrastive loss to a(n) (variational)autoencoder
- Getting started with real robots
Topic 1: Discuss idea for adding contrastive loss to a(n) (variational) autoencoder
- Test different architectures
- Compare the proposed architecture(contrastive AE) with normal contrastive or (V)AEs
- (future) integrateprior knowledge
- Explore different ways to update the weights (gradient flow)
- contrastive loss flows or not through the decoder
- check the learning stability
- we need stable update rules
- Derive the equations from Information Theoretic Principles
- e.g. ELBO
Topic 2: Getting started with real robots
- Talk with Vedant and join his project
- Do the FE tutorials on Gazebo
Next Steps
- Derive the Information Theoretic equations
- Start building AEs and Contrastive loss networks
- Do the FE Panda tutorials
Literature
[1] Srinivas, Aravind, Michael Laskin, and Pieter Abbeel. “Curl: Contrastive
unsupervised representations for reinforcement learning.” arXiv preprint arXiv:2004.04136 (2020).
Next Meeting
TBA