Reproduce “On the Robustness of Monte Carlo Dropout Trained with Noisy Labels” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Purvi Goel, Li Chen
Researchers at Facebook investigated the robustness of Monte Carlo Dropout (MCDropout) against noisy labels, demonstrating its consistent superior performance compared to deterministic models. Their analysis revealed that MCDropout fosters latent representations characterized by reduced neuron volatility and increased network sparsity, which contributes to better generalization.
Key insights
- MCDropout effectively delays the memorization of noisy labels, leading to improved generalization on clean test data.
- The robustness of MCDropout is attributed to its influence on latent representations, specifically by inducing reduced neuron volatility and increased network sparsity.
- MCDropout acts as a strong regularizer by limiting neuron activation and responsiveness, thereby constraining the model's capacity to overfit to corrupted training data.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.