Pixel Club: On The Resistance of Neural Networks to Label Noise

Speaker:
Amnon Drory (Tel-Aviv University)
Date:
Tuesday, 22.1.2019, 11:30
Place:
Room 337 Taub Bld.

Neural Networks have been shown to be remarkably resistant to label noise. This means that you can train a network using a data set that contains a significant fraction of wrongly-labelled samples, and the network will still be able to accurately predict the label for a previously unseen sample. Our results show that a main factor in explaining this resistance is that networks learn from a local group of training samples, similarly to a K-Nearest Neighbors algorithm. We are able to provide a mathematical expression for the expected accuracy of a network given the noise level, for a general family of noise types. We also show that the extent of the resistance depends strongly on how localized the noise is.

Shared work with: Oria Ratzon, Prof. Shai Avidan and Dr. Raja Giryes at Tel-Aviv University, school of Electrical Engineering.

Back to the index of events