Estimating ReLU Deep Neural Networks Uncertainty Estimations
Name
Toomas Roosma
Abstract
In this thesis, we investigated the ability of ReLU neural networks to predict its uncertainty estimations. Knowing the uncertainty of predictions is essential because it helps to understand the model’s reliability. Four different experiments were conducted, and the accuracy of neural network uncertainty estimations was measured in each. The thesis found that both the reciprocal of the logarithm of the number of training points in a region and the number of breakpoints affect predictions. The closer the training data and breakpoint are located to the predicted value; the more importance they have.
Graduation Thesis language
Estonian
Graduation Thesis type
Bachelor - Computer Science
Supervisor(s)
Meelis Kull
Defence year
2023