Learning in an Uncertain World: Representing Ambiguity Through Multiple\n Hypotheses

2016/12/01 by Christian Rupprecht, Iro Laina, Rupprecht, Christian +11 · 10 citations
Computer Science · #Anomaly Detection Techniques and Applications #Computer Vision and Pattern Recognition (cs.CV) #Explainable Artificial Intelligence (XAI) #FOS: Computer and information sciences #Human Pose and Action Recognition #Machine Learning and Data Classification

paper · pdf · doi:10.48550/arxiv.1612.00197

openalex publication_date 2016/12/01 · openalex created_date 2022/10/05 · openalex updated_date 2026/07/28

Abstract

Many prediction tasks contain uncertainty. In some cases, uncertainty is\ninherent in the task itself. In future prediction, for example, many distinct\noutcomes are equally valid. In other cases, uncertainty arises from the way\ndata is labeled. For example, in object detection, many objects of interest\noften go unlabeled, and in human pose estimation, occluded joints are often\nlabeled with ambiguous values. In this work we focus on a principled approach\nfor handling such scenarios. In particular, we propose a framework for\nreformulating existing single-prediction models as multiple hypothesis\nprediction (MHP) models and an associated meta loss and optimization procedure\nto train them. To demonstrate our approach, we consider four diverse\napplications: human pose estimation, future prediction, image classification\nand segmentation. We find that MHP models outperform their single-hypothesis\ncounterparts in all cases, and that MHP models simultaneously expose valuable\ninsights into the variability of predictions.\n

Citations

• Very Deep Convolutional Networks for Large-Scale Image Recognition
• The Pascal Visual Object Classes (VOC) Challenge
• Centroidal Voronoi Tessellations: Applications and Algorithms
• Gradient-based learning applied to document recognition
• ImageNet Large Scale Visual Recognition Challenge
• Least squares quantization in PCM
• CNN-RNN: A Unified Framework for Multi-label Image Classification
• NTU RGB+D: A Large Scale Dataset for 3D Human Activity Analysis
• Faster R-CNN: Towards Real-Time Object Detection with Region Proposal\n Networks
• Exploit Bounding Box Annotations for Multi-label Object Recognition
• Fully Convolutional Networks for Semantic Segmentation
• Dropout as a Bayesian Approximation: Representing Model Uncertainty in Deep Learning

Cited by

• TimePre: Bridging Accuracy, Efficiency, and Stability in Probabilistic Time-Series Forecasting
• Controlling Contrastive Self-Supervised Learning with Knowledge-Driven Multiple Hypothesis: Application to Beat Tracking
• S3OD: Towards Generalizable Salient Object Detection with Synthetic Data
• Test-time Uncertainty Estimation for Medical Image Registration via Transformation Equivariance
• From Samples to Scenarios: A New Paradigm for Probabilistic Forecasting
• Explorative Modeling: Unlocking a Third Pretraining Axis and End-to-End Generation
• Label-Efficient Grasp Joint Prediction with Point-JEPA
• Structured Basis Function Networks: Loss-Centric Multi-Hypothesis Ensembles with Controllable Diversity
• Prior-Guided Residual Diffusion: Calibrated and Efficient Medical Image Segmentation
• MATPAC++: Enhanced Masked Latent Prediction for Self-Supervised Audio Representation Learning

Related

• A Tale Of Two Long Tails
• Learning from Uncertain Data: From Possible Worlds to Possible Models
• Can You Trust Your Model's Uncertainty? Evaluating Predictive Uncertainty Under Dataset Shift
• Distribution Preserving Multiple Hypotheses Prediction for Uncertainty Modeling
• Training Uncertainty-Aware Classifiers with Conformalized Deep Learning