Please enable JavaScript.
Coggle requires JavaScript to display documents.
Literature (Uncertainty Methods (Temperature Scaling, Dropout:…
Literature
-
Dropout: Reinterpreting Dropout as Variational Inference
-
-
-
-
-
Types of Uncertainty
Aleatoric Uncertainty: inherent ambiguity in outputs for a given input (more)
Heteroscedastic: depends on input data, is predicted as model output
Homoscedastic: does not depend on input data, not a model output but rather a quantity varying between tasks
Epistemic Uncertainty: Parameter uncertainty, reducable through additional data
-
(Measurement Uncertainty): not so common, noisy measurement system
-
Current Challenges
Real-Time capability, best without Monte Carlo Inference
Benchmark, like ImageNet for Computer Vision
-
-
Approximators
Bayesian Approximators
Variational Inference (1) and (2)
Dropout based Variational Inference (1) and (2)
-
-
-
-
-
-
-
-
Code
Experiment
-
-
-
Attach Prior and Posterior from Agnieszka to Network, Try to produce LU, EU, AU
Modes
- Varying the amount of data
- Noise in the data / no noise
-
Architecture
Framework from Probabilistic U-Net, just DenseNet
-
-