Professional Documents
Culture Documents
Project Progress
Project Progress
Published in 2020
Publisher: Springer International Publishing;Springer, Year: 2020
Presented by Abdullah
Reg no 2273
Subject
Probability and stochastic
Data Augmentation
• Data augmentation is an effective and universal technique for
improving generalization performance of deep neural networks.
• To augment data we use different techniques
• U-ResNet
• nnUNet NoDA
• nnUNet
• SCNAS
Proposed Techniques
• we formulate the problem of finding the optimal augmentation policy
as a discrete search problem.
• For this purpose we use a techniques:
• search space
• search algorithm
Comparison in term of lost
Search space
• First work for applying Auto Augment strategies in medical image
area.
• So to solve the problem we design:
• search space for our ASNG algorithm
• Consist of seven image operation to be applied on image sequentially.
• Each image operation is associated with two hyper parameters
• probability of applying this operation
• the interval of magnitude for the image operation
Stochastic Relaxation
• We turn the original optimization problem into an optimization of
differentiable objective J by stochastic relaxation.
• Instead of directly optimizing w.r.t c,
• we consider a distribution over c parametrized by θ,
• and minimize the expected value of the validation loss w.r.t θ