Discretization-Induced Dirichlet Posterior for Robust Uncertainty Quantification on Regression
Authors: Xuanlong Yu, Gianni Franchi, Jindong Gu, Emanuel Aldea
AAAI 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Extensive experiments on age estimation, monocular depth estimation, and super-resolution tasks show that our proposed method can provide robust uncertainty estimates in the face of noisy inputs and that it can be scalable to both image-level and pixel-wise tasks. 4 Experiments In this section, we first show the feasibility of the proposed generalized Aux UE on toy examples. Then, we demonstrate the effectiveness of epistemic uncertainty estimation using the proposed DIDO on age estimation and monocular depth estimation (MDE) tasks, and investigate the robustness of aleatoric uncertainty estimation on MDE task. |
| Researcher Affiliation | Academia | Xuanlong Yu1,2, Gianni Franchi2, Jindong Gu3, Emanuel Aldea1 1SATIE, Paris-Saclay University 2U2IS, ENSTA Paris, Institut Polytechnique de Paris 3University of Oxford |
| Pseudocode | No | The paper does not contain structured pseudocode or algorithm blocks. |
| Open Source Code | Yes | Details and demo-code can be found in Supp Section C.1 and C.2 respectively. |
| Open Datasets | Yes | We train the models on AFAD (Niu et al. 2016) training set and choose AFAD test set as the ID dataset for the OOD detection task. We take CIFAR10 (Krizhevsky, Hinton et al. 2009), SVHN (Netzer et al. 2011), MNIST (Le Cun 1998), Fashion MNIST (Xiao, Rasul, and Vollgraf 2017), Oxford-Pets (Parkhi et al. 2012) and Noise image generated by Pytorch (Paszke et al. 2019) (Fake Data) as the OOD datasets. ... We use BTS (Lee et al. 2019) as the main task model and KITTI (Geiger et al. 2013; Uhrig et al. 2017) Eigensplit (Eigen, Puhrsch, and Fergus 2014) training set for training both BTS and Aux UE models. |
| Dataset Splits | Yes | We generate KITTI-C from KITTI Eigen-split validation set using the code of Image Net-C (Hendrycks and Dietterich 2019)... We take all the valid pixels from the KITTI validation set (ID) as the negative samples and the valid pixels from the NYU (Nathan Silberman and Fergus 2012) validation set (OOD) as the positive samples. |
| Hardware Specification | No | The paper does not provide specific hardware details (e.g., exact GPU/CPU models, memory amounts, or detailed computer specifications) used for running its experiments, only mentioning general terms like 'DNNs'. |
| Software Dependencies | No | Noise image generated by Pytorch (Paszke et al. 2019) |
| Experiment Setup | Yes | We define a main task DNN fω with trainable parameters ω as shown in the blue area in Fig. 1. ... we set the initial α as 1 so that the Dirichlet concentration parameters can be formed as in (Sensoy, Kaplan, and Kandemir 2018; Charpentier, Z ugner, and G unnemann 2020): α(i) = e(i) + 1 = σΘ2(x(i)) + 1, where e(i) is given by an exponential function on the top of σΘ2. Then we minimize the Kullback Leibler (KL) divergence between the variational distribution P(π|x, Θ2) and the true posterior P(π| D) to achieve bΘ2: ... where ψ is the digamma function, λ is a positive hyperparameter for the regularization term and ϵ is given by Eq. 2. |