How far can you trust your Deep Neural Networks? (feat. TensorFlow) (25min)

• Science / Data
• Experienced
• 2017-08-13 (Sun) 16:00 - 16:40
• English
• 201
• Sungjoon Choi / sungjoonchoi
• Photography and recording is not allowed

Slides

https://goo.gl/iJQ7JG

Description

In this talk, I will mainly focus on inferring the uncertainty information when using a deep neural network in TensorFlow. In particular, regression tasks, which have been less focussed compared to classification problems, will be mainly considered. First, a mixture density network will be implemented with TensorFlow where its superiority will be shown compared to ordinary regression networks. Then, two different methods, epistemic uncertainty and the entropy of a Gaussian mixture model, will be presented to estimate the uncertainty information along with the prediction output.

Comments

blog comments powered by Disqus