Correlated Parameters to Accurately Measure Uncertainty in Deep Neural Networks

Posch, Konstantin; Pilz, Jürgen

doi:10.1109/tnnls.2020.2980004

Cited by 32 publications

(15 citation statements)

References 33 publications

Supporting

Mentioning

Contrasting

Unclassified

Order By: Relevance

“…Thus, the network weights and biases are also normally distributed,

The diagonal covariance matrix of the Gaussian variational distributions assumes the independence of all the network parameters and biases. The calculation of the respective gradients and an extension to a tridiagonal covariance matrix can be found in [ 44 ] and [ 52 ]. We implement the Bayesian segmentation network using the open source Python library PyTorch [ 53 ].…”

Section: Methodsmentioning

confidence: 99%

From a Point Cloud to a Simulation Model—Bayesian Segmentation and Entropy Based Uncertainty Estimation for 3D Modelling

Petschnigg

Spitzner

Weitzendorf

et al. 2021

Entropy

Self Cite

View full text Add to dashboard Cite

The 3D modelling of indoor environments and the generation of process simulations play an important role in factory and assembly planning. In brownfield planning cases, existing data are often outdated and incomplete especially for older plants, which were mostly planned in 2D. Thus, current environment models cannot be generated directly on the basis of existing data and a holistic approach on how to build such a factory model in a highly automated fashion is mostly non-existent. Major steps in generating an environment model of a production plant include data collection, data pre-processing and object identification as well as pose estimation. In this work, we elaborate on a methodical modelling approach, which starts with the digitalization of large-scale indoor environments and ends with the generation of a static environment or simulation model. The object identification step is realized using a Bayesian neural network capable of point cloud segmentation. We elaborate on the impact of the uncertainty information estimated by a Bayesian segmentation framework on the accuracy of the generated environment model. The steps of data collection and point cloud segmentation as well as the resulting model accuracy are evaluated on a real-world data set collected at the assembly line of a large-scale automotive production plant. The Bayesian segmentation network clearly surpasses the performance of the frequentist baseline and allows us to considerably increase the accuracy of the model placement in a simulation scene.

show abstract

“…Thus, the network weights and biases are also normally distributed,

Section: Methodsmentioning

confidence: 99%

From a Point Cloud to a Simulation Model—Bayesian Segmentation and Entropy Based Uncertainty Estimation for 3D Modelling

Petschnigg

Spitzner

Weitzendorf

et al. 2021

Entropy

Self Cite

View full text Add to dashboard Cite

show abstract

“…In most of the previous studies, researchers in UQ are splitting the dataset randomly 10,11 . The performance of the trained NN varies based on the data splitting 12,13 . Researchers often randomly split data several times and then train NNs for each split and usually get average performance.…”

Section: Background and Summarymentioning

confidence: 99%

Synthetic Datasets for Numeric Uncertainty Quantification

Kabir

Abdar

Khosravi

et al. 2021

Preprint

View full text Add to dashboard Cite

In this paper, we propose ten synthetic datasets for point prediction and numeric uncertainty quantification (UQ). These datasets are split into train, validation, and test sets for model benchmarking. Equations and the description of each dataset are provided in detail. We also present representative shallow neural network (NN) training and Random Vector Functional Link (RVFL) training examples. Both training train models for the point prediction. We perform uncertainty quantification with the consideration of a gaussian and homoscedastic distribution. As the distribution consideration and models are rudimental, much room exists for further explorations and improvements. The dataset and scripts are available at the following link: https://github.com/dipuk0506/UQ-Data

show abstract

“…The following subsections present in more details about the most common and recent calibration techniques (temperature scaling [27], Monte Carlo Dropout [29], [42]) and regularization techniques (penalization of overconfidence output distributions [24], [26], [28], [39], [41], label smoothing [43]). Additionally, we would discuss the disadvantages of the mentioned techniques when predicting objects belonging to out-of-training-distribution data (which may be critical in autonomous driving and robotics).…”

Section: Related Work On Overconfidence Predictionmentioning

confidence: 99%

Probabilistic Approach for Road-Users Detection

Melotti¹,

Lu²,

Conde³

et al. 2021

Preprint

View full text Add to dashboard Cite

Object detection in autonomous driving applications implies that the detection and tracking of semantic objects are commonly native to urban driving environments, as pedestrians and vehicles. One of the major challenges in state-of-the-art deep-learning based object detection is false positive which occurrences with overconfident scores. This is highly undesirable in autonomous driving and other critical roboticperception domains because of safety concerns. This paper proposes an approach to alleviate the problem of overconfident predictions by introducing a novel probabilistic layer to deep object detection networks in testing. The suggested approach avoids the traditional Sigmoid or Softmax prediction layer which often produces overconfident predictions. It is demonstrated that the proposed technique reduces overconfidence in the false positives without degrading the performance on the true positives. The approach is validated on the 2D-KITTI objection detection through the YOLOV4 and SECOND (Lidar-based detector). The proposed approach enables enabling interpretable probabilistic predictions without the requirement of re-training the network and therefore is very practical.

show abstract

Correlated Parameters to Accurately Measure Uncertainty in Deep Neural Networks

Cited by 32 publications

References 33 publications

From a Point Cloud to a Simulation Model—Bayesian Segmentation and Entropy Based Uncertainty Estimation for 3D Modelling

From a Point Cloud to a Simulation Model—Bayesian Segmentation and Entropy Based Uncertainty Estimation for 3D Modelling

Synthetic Datasets for Numeric Uncertainty Quantification

Probabilistic Approach for Road-Users Detection

Contact Info

Product

Resources

About