Why can deep ensembles improve uncertainty estimates?

Question: What is the most accurate explanation for why independently trained deep ensembles can improve predictive uncertainty? A) Their averaged output is mathematically guaranteed to equal the exact Bayesian posterior predictive distribution B) Different initializations and training trajectories can produce diverse predictive functions whose disagreement carries information about model uncertainty C) Averaging eliminates aleatoric uncertainty while retaining only epistemic uncertainty D) Ensemble variance is guaranteed to be calibrated even when every member is trained on identical corrupted labels Correct: B Explanation: Independently trained networks can converge to different solutions and make different predictions away from well-supported regions. Their disagreement can provide useful uncertainty information, although the method is not an exact Bayesian posterior and does not automatically guarantee calibration. Topic: advanced ML / uncertainty estimation / deep ensembles