back to list

Project: Explaining learned features by generative models


In order to get some insight into the inner workings of deep neural network classifiers, a method that enables the interpretation of learned features would be very helpful. This master project is loosely based on the approach presented in [1], where a GAN is learned to generate examples that are close to the decision boundary of a classifier. A similar approach could be used to generate examples where one learned feature is particularly highly expressed. Learned features are e.g., the output of the penultimate layer in a deep neural network. 

This topic might be for you if you like to explore your own ideas on how to make such a generator happen. Ideally, you have trained a generative model before and know about pitfalls and the impact of design choices of the model and the loss function.

[1] Cunha, Luís, et al. "GASTeN: Generative Adversarial Stress Test Networks." International Symposium on Intelligent Data Analysis., 2023.

Sibylle Hess
Secondary supervisor
Wil Michiels
Get in contact