<ul><li>Researchers propose a Compressed Data Representation Model (CDRM) to estimate aleatoric and epistemic uncertainty in a learned regressive system dynamics model.</li><li>CDRM disentangles the inherent randomness of the system (aleatoric uncertainty) from the lack of data (epistemic uncertainty) to enable risk-aware control, reinforcement learning, efficient exploration, and robust policy transfer.</li><li>CDRM incorporates a neural network encoding of the data distribution and an inference procedure based on Langevin dynamics sampling for predicting arbitrary output distributions.</li><li>Empirical evaluations demonstrate that CDRM outperforms existing methods in identifying aleatoric and epistemic uncertainties separately, even in datasets with multimodal output distributions.</li></ul>

Disentangling Uncertainties by Learning Compressed Data Representation

Discover more