# Deep Deterministic Uncertainity
[[Uncertainty in Machine Learning]] can be levaraged as data-acquisition function for active learning.
Aleatoric uncertainity (data uncertainity)
- Ambiguous between multiple classes (i.e. data with high density but high entropy)
- Doesn't decrease with more data.
Epistemic uncertainity (model uncertainity)
- Out-of-distribution (i.e. data point with low density)
- Model unsure because it hasn't seen enough of similar data, decreases with more data.
- Can be estimated by fitting a density function p(X) over data, in this case a [[Gaussian Mixture Model]]
Goal is to acquire labels for the most efficient samples. - Label data with high epistemic uncertainity as that helps the model the most.
- Labeling data with high aleatoric uncertainity isn't that helpful.
![[Deep Deterministic Uncertainity Algotirhm.png]]
---
## References