Multitask Optimization Complexity
Aravind explains that the contrast loss behaves like a classification task, playing well with reinforcement learning. However, predicting the future in high-dimensional spaces can dominate the RL loss, requiring careful hyperparameter tuning for balance. The contrast loss, resembling supervised learning, offers a smoother path for multi-objective optimization.In this clip
From this podcast
Machine Learning Street Talk (MLST)
CURL: Contrastive Unsupervised Representations for Reinforcement Learning
Related Questions