PhD candidate, Robustness Verification for Meta-learned Neural Networks

Updated: over 2 years ago
Deadline: 30 Nov 2021

The Faculty of Science and Liacs Institute of Advanced Computer Science is looking for a

PhD candidate, Robustness Verification for Meta-learned Neural Networks

Vacancy number 21-454

Key responsibilities

Neural networks achieve state-of-the-art performance on many image-recognition tasks. Despite this enormous potential, it is widely acknowledged that neural networks also need large amounts of data and high GPU requirements to achieve this performance. Also, neural networks can be subject to adversarial attacks. The data- and GPU-related limitations can be addressed by meta-learning techniques, where such a neural network is pre-trained on similar tasks. This gives rise to few-shot learning, where neural networks have shown to be competitive if as few as five examples can be provided of a given class. The possibility of an adversarial attack is often addressed by neural network verification techniques that certify the robustness of a neural network. This Ph.D. trajectory will work on the intersection of meta-learning and neural network verification. 

Goal: The successful candidate will carry out research on the intersection of neural network verification and meta-learning. In particular, they will identify, construct and evaluate possible benchmarks that can be used across the research community. Additionally, we will develop techniques that combine the fundamental concepts and techniques from neural network verification and meta-learning. By applying AutoML techniques (e.g., Bayesian optimization, bandit methods) on meta-learning methods, we can search for hyperparameter configurations that do not only optimize for performance, but also take into account the robustness of a neural network.

Embedding: This project is part of the TAILOR network (https://tailor-network.eu/), a collaborative project containing the top research labs and industry partners across Europe (members from, e.g., Leiden University, University of Freiburg, INRIA). The candidate will be working with Prof. Dr. Holger Hoos and Dr. Jan N. van Rijn, from the Automated Design of Algorithms Research Group (ada.liacs.nl, LIACS, Leiden University). Our group has deep and broad expertise in all areas of machine learning relevant to this project and plays a key role in a large vibrant international research network; this will provide the candidate with ample opportunity to start a successful scientific career in one of the hottest areas of artificial intelligence.



Similar Positions