University of Cambridge > > Machine Learning @ CUED > Semi-Unsupervised Learning with Deep Generative Models / Disentangling Improves VAEs' Robustness to Adversarial Attacks

Semi-Unsupervised Learning with Deep Generative Models / Disentangling Improves VAEs' Robustness to Adversarial Attacks

Add to your list(s) Download to your calendar using vCal

If you have a question about this talk, please contact Eric T Nalisnick.

This talk will consist of two parts (approx 20 min each)

#1 Semi-Unsupervised Learning with Deep Generative Models

Semi-Unsupervised learning is a new regime that may be quite common in the real world. It is a form of ultra-sparse semi-supervised learning, where for some classes in our data there are no labelled examples in the training set at all, only unlabelled examples. It could be due to selection biases in how we obtained our annotated data compared to our larger amount of unlabelled data. Or it could just be that we were unlucky and missed some classes in the small labelled dataset.

This type of data motivates models that can jointly perform clustering and semi-supervised learning. After training we want to have a classifier that accurately predicts both semi-supervised classes and classes for which it has never seen a single labelled example.

It is well known that some DGMs developed to learn a classifier in the semi-supervised regime can fail when trained on unlabelled data only—-they simply don’t learn to discern between classes of data. We cover why this happens and present some models we have developed that enable us to learn in this set-up, and show some results.

#2 Disentangling Improves VAEs’ Robustness to Adversarial Attacks

We highlight that conventional VAEs are brittle under attacks that aim to fool them into reconstructing a subtly-distorted input to a chosen target image.

However, methods recently introduced for disentanglement such as β-TCVAE (Chen et al., 2018) improve robustness to proposed adversarial attacks that aim to match the distorted input to the target in a VAE ’s latent space. This motivated us to develop Seatbelt-VAE, a new hierarchical disentangled VAE that is designed to be significantly more robust to adversarial attacks than existing approaches, while retaining high quality reconstructions.

This talk is part of the Machine Learning @ CUED series.

Tell a friend about this talk:

This talk is included in these lists:

Note that ex-directory lists are not shown.


© 2006-2024, University of Cambridge. Contact Us | Help and Documentation | Privacy and Publicity