Towards Interpretability of Segmentation Networks by analyzing DeepDreams

Abstract :

Interpretability of a neural network can be expressed as the identification of patterns or features to which the network can be either sensitive or indifferent. To this aim, a method inspired by DeepDream is proposed, where the activation of a neuron is maximized by performing gradient ascent on an input image. The method outputs curves that show the evolution of features during the maximization. A controlled experiment show how it enables assess the robustness to a given feature, or by contrast its sensitivity. The method is illustrated on the task of segmenting tumors in liver CT images.

Complete list of metadatas

https://hal.telecom-paristech.fr/hal-02288076
Contributor : Telecomparis Hal <>
Submitted on : Friday, September 13, 2019 - 5:39:19 PM
Last modification on : Thursday, October 17, 2019 - 12:37:00 PM

Identifiers

  • HAL Id : hal-02288076, version 1

Citation

Vincent Couteaux, O. Nempont, Guillaume Pizaine, Isabelle Bloch. Towards Interpretability of Segmentation Networks by analyzing DeepDreams. iMIMIC Workshop at MICCAI 2019: Interpretability of Machine Intelligence in Medical Image Computing, 2019, Shenzhen, China. ⟨hal-02288076⟩

Share

Metrics

Record views

29