Next week's seminar will feature Yuhao Zhang, who will be presenting on "Future Directions in Certifiable Machine Learning", you can read more in the abstract below! Like always, it will be at 1:05pm in room 3310, or on
zoom.
Abstract:
"Certified
deep learning has emerged as an important research area that seeks to provide provably correct and reliable deep learning algorithms. However, relying solely on certification can create a misleading sense of safety, which is a major challenge in this field.
In this presentation, we will explore the topic of bridging the gap between certified and explainable deep learning. In
addition to bridging the gap between certified and explainable deep learning, we also discuss the usage of generative models in randomized smoothing (RS). Generative models have been shown to improve the certified accuracy in RS. However, there is a concern
that pretrained generative models could be attacked, leading to unsafe RS pipelines. Finally, we discuss the
topic of proof sharing for an ensemble of neural networks. Ensemble methods are commonly used to improve the accuracy and robustness of deep learning models. However, verifying the correctness of an ensemble of similar neural networks can be computationally
expensive. We propose to share some intermediary proofs between the individual models in the ensemble, which can significantly reduce the computational complexity of the verification process."
Hope you can attend,
Lauren Neudorf
Program Manager, MadPL Research Group
University of Wisconsin-Madison
Department of Computer Sciencesâ
(716) 704-4463
(she/her/hers)