[논문 리뷰] What Uncertainties Do We Need in Bayesian Deep Learning for Computer Vision?

Alex Kendall, Yarin Gal|arXiv (Cornell University)|2017. 03. 15.

Domain Adaptation and Few-Shot Learning참고 문헌 30인용 수 2,528

한 줄 요약

본 논문은 입력 의존적 해상도 불확실성과 인지적 불확실성을 결합하는 통합 베이지안 프레임워크를 개발하여 픽셀 단위 깊이 추정 및 의미론적 분할을 개선하고, 학습된 불확실성을 통한 손실 감소를 도입한다.

ABSTRACT

There are two major types of uncertainty one can model. Aleatoric uncertainty captures noise inherent in the observations. On the other hand, epistemic uncertainty accounts for uncertainty in the model -- uncertainty which can be explained away given enough data. Traditionally it has been difficult to model epistemic uncertainty in computer vision, but with new Bayesian deep learning tools this is now possible. We study the benefits of modeling epistemic vs. aleatoric uncertainty in Bayesian deep learning models for vision tasks. For this we present a Bayesian deep learning framework combining input-dependent aleatoric uncertainty together with epistemic uncertainty. We study models under the framework with per-pixel semantic segmentation and depth regression tasks. Further, our explicit uncertainty formulation leads to new loss functions for these tasks, which can be interpreted as learned attenuation. This makes the loss more robust to noisy data, also giving new state-of-the-art results on segmentation and depth regression benchmarks.

연구 동기 및 목표

• Differentiate and quantify aleatoric versus epistemic uncertainty in vision models.
• Develop a framework that combines input-dependent aleatoric uncertainty with epistemic uncertainty.
• Demonstrate robustness to noisy data via learned loss attenuation.
• Apply the framework to per-pixel semantic segmentation and depth regression tasks.
• Show practical trade-offs between uncertainty types and inference efficiency.

제안 방법

• Model epistemic uncertainty with dropout-based approximate Bayesian inference (Monte Carlo dropout).
• Model aleatoric uncertainty by learning per-pixel observation noise (sigma) and its log-variance s via a stable loss that uses exp(-s).
• Combine aleatoric and epistemic uncertainties in a single Bayesian CNN with a shared head predicting both y and sigma^2.
• Extend heteroscedastic uncertainty to classification by marginalizing logit-space Gaussian noise with Monte Carlo estimation (softmax over perturbed logits).
• Derive a loss L_BNN = (1/D) sum_i 0.5 exp(-s_i) ||y_i - y_hat_i||^2 + 0.5 s_i for regression, and a corresponding stochastic cross-entropy variant for classification.
• Assess uncertainty quality via precision-recall and calibration plots; analyze behavior with training data size and out-of-distribution tests.

실험 결과

연구 질문

• RQ1 Can modeling aleatoric uncertainty alone, epistemic uncertainty alone, or their combination improve vision tasks like segmentation and depth regression?
• RQ2 Does learned loss attenuation from aleatoric uncertainty improve robustness to noisy labels and data?
• RQ3 How do aleatoric and epistemic uncertainties differ in their behavior with respect to data distance from the training distribution and dataset size?
• RQ4 Is modeling both uncertainties compatible with real-time or near-real-time inference?
• RQ5 What are the empirical gains on standard benchmarks for segmentation and depth regression when combining uncertainties?

주요 결과

• 모델링된 aleatoric 및 epistemic 불확실성은 CamVid 및 NYUv2 분할과 Make3D 및 NYUv2 깊이 회귀에서 최첨단 성능을 제공합니다.
• CamVid 및 NYUv2에서 분할에 대해 aleatoric 불확실성이 epistemic 불확실성보다 더 큰 이점을 제공하는 반면, 두 가지를 결합하면 추가적인 개선이 나타납니다.
• 두 불확실성의 결합으로 깊이 회귀에서 이득이 증가하며 Make3D와 NYUv2 데이터셋 전반에 걸쳐 상대적 개선이 관찰됩니다.
• 알고 있는 손실 완화로 인해 학습된 aleatoric 불확실성은 어렵거나 잘못 표기된 픽셀을 낮춰 다소 robust하게 만들어 주며, 전체 성능을 해치지 않습니다.
• 인지적 불확실성은 더 많은 학습 데이터와 함께 감소하며, out-of-distribution 또는 데이터가 적은 시나리오에서 더 두드러집니다; 반면 aleatoric 불확실성은 데이터 크기에 상대적으로 안정적입니다.
• 모델은 aleatoric 구성요소에 대해 실시간에 가까운 추론을 달성하는 반면, 몬테카를로 드롭아웃을 통한 전체 인지적 샘플링은 상당한 계산 비용이 필요하여 실용적 배치에 대한 trade-off를 강조합니다.