REVIEW 8 cited by
Deep Deterministic Uncertainty: A Simple Baseline
Not yet reviewed by Pith; the record is open.
This paper has not been read by Pith yet. Machine review is queued; the pith claim, tier, and objections will appear here once it completes.
SPECIMEN: schema-true, not a live event
T0 review · schema-true
One-sentence machine reading of the paper's core claim.
pith:XXXXXXXX · record.json · timestamp
Signed reviews
read the original abstract
Reliable uncertainty from deterministic single-forward pass models is sought after because conventional methods of uncertainty quantification are computationally expensive. We take two complex single-forward-pass uncertainty approaches, DUQ and SNGP, and examine whether they mainly rely on a well-regularized feature space. Crucially, without using their more complex methods for estimating uncertainty, a single softmax neural net with such a feature-space, achieved via residual connections and spectral normalization, *outperforms* DUQ and SNGP's epistemic uncertainty predictions using simple Gaussian Discriminant Analysis *post-training* as a separate feature-space density estimator -- without fine-tuning on OoD data, feature ensembling, or input pre-procressing. This conceptually simple *Deep Deterministic Uncertainty (DDU)* baseline can also be used to disentangle aleatoric and epistemic uncertainty and performs as well as Deep Ensembles, the state-of-the art for uncertainty prediction, on several OoD benchmarks (CIFAR-10/100 vs SVHN/Tiny-ImageNet, ImageNet vs ImageNet-O) as well as in active learning settings across different model architectures, yet is *computationally cheaper*.
Forward citations
Cited by 8 Pith papers
-
Constrained Co-Design for Photonic Bayesian Neural Networks
Photonic BNN inference is a constrained variational problem; most hardware limits are trainable through, but signed mean range and some scale bounds are hard representational limits that require hardware changes.
-
A Unified Evaluation Framework for Epistemic Predictions
A new metric E = KL distance to the nearest credal-set vertex plus λ times non-specificity ranks uncertainty-aware classifiers under a user-chosen accuracy-precision trade-off.
-
Density-Informed Pseudo-Counts for Calibrated Evidential Deep Learning
DIP-EDL sets Dirichlet pseudo-counts to the product of marginal input density and learned class probabilities, concentrating on the true label distribution while sending OOD inputs to the prior.
-
Optimizing $pp\to A\to Z^{*}h\to \ell^+\ell^- b\bar b$ Searches at the LHC in the 2HDM Type-I with Inverted Hierarchy
In 2HDM Type-I with inverted hierarchy, applying m_ll in 20-50 GeV and m_bb below 100 GeV makes A to Z* h to ll bb stand out against Drell-Yan and ttbar backgrounds.
-
HQFNN: A Compact Quantum-Fuzzy Neural Network for Accurate Image Classification
HQFNN is a quantum-fuzzy neural network that embeds fuzzy membership and defuzzification in a small simulated quantum circuit, and it is reported to outperform several specialized baselines on five small image datasets.
-
Uncertainty and Energy based Loss Guided Semi-Supervised Semantic Segmentation
A semi-supervised semantic segmentation method combining aleatoric uncertainty loss and energy-based loss with union-intersection pseudo-labels gives modest gains over the CPCL baseline on PASCAL VOC and Cityscapes, b...
-
Probabilistic Modeling of Disparity Uncertainty for Robust and Efficient Stereo Matching
Stereo matching can report separate data and model uncertainty by combining ordinal-regression disparity distributions with a kernel-regression model-uncertainty estimator.
-
Risk-aware Classification via Uncertainty Quantification
Risk-aware evidential classifiers that learn a per-sample prior from a misclassification cost matrix lower expected cost on MNIST and CIFAR10 versus cost-sensitive baselines.
Discussion (0). Continue with ORCID to comment.