2 full-time PhD positions in Robust Learning of Sparse Representations

Organisation

Founded in 1614, the University of Groningen enjoys an international reputation as a dynamic and innovative institution of higher education offering high-quality teaching and research. Flexible study programmes and academic career opportunities in a wide variety of disciplines encourage the 36,000 students and researchers alike to develop their own individual talents. As one of the best research universities in Europe, the University of Groningen has joined forces with other top universities and networks worldwide to become a truly global centre of knowledge.

The two PPs will be developed by two paid full-time PhD projects at the Bernoulli Institute for Mathematics, Computer Science, and Artificial Intelligence, Faculty of Science and Engineering, University of Groningen. The candidates will also be affiliated with the CogniGron Center, which is funding the projects.

Within the Faculty of Science and Engineering, two 4-year PhD positions are available at the Bernoulli Institute for Mathematics, Computer Science and Artificial Intelligence with the topic of Robust Learning of Sparse Representations: Brain-inspired Inhibition and Statistical Physics Analysis. Both candidates would become members of the Computer Science Department, with the candidate of PP1 embedded in the group Information Systems and the other candidate of PP2 embedded in the group Intelligent Systems.

Job description

Both positions will be supported by the Groningen Cognitive systems and Materials Center (CogniGron), see

https://www.rug.nl/research/fse/cognitive-systems-and-materials/

The high energy efficiency and generalization abilities in processing sensory information achieved by a mammalian brain are among the gold nuggets of neuromorphic computing. In contrast to conventional super/cluster/grid computing centers which take very large spaces and consume a great deal of energy, a human brain weighs less than 1500g and requires only 20w of power to operate. Emulating these remarkable abilities of the brain represents the pursuit of neuromorphic computing in achieving more with less.

Sparsity is among the key factors that contribute to high energy efficient processing in the brain. Neuroscientists believe that inhibition is a crucial property that results in sparse and thus highly energy efficient representations. Sparsity and inhibition are the focus of this joint project, which consists of two key objectives formulated in two complementary PhD Projects (PPs).

The PPs are complementary to each other in that they contribute to the repeating pair of the core building blocks in deep architectures of deep architectures, namely convolutions (PP1) and activation functions (PP2). Understanding the interplay and integration of these two concepts in one system will be targeted in joint efforts of the two PPs throughout the project.

In this joint project we will:

1. PhD Project 1 (PP1). Daily supervisor: Dr. George Azzpardi; Co-supervisors: Prof. Michael Biehl and Prof. Dimka Karastoyanova; External co-supervisor: Dr. Nicola Strisciuglio (University of Twente)

• investigate the best trade-off between effectiveness and high-efficient energy inhibition-augmented networks, and
• develop methodologies for the embedding of the push-pull component in state-of-the-art convolutional neural networks (CNNs) and spiking neural networks (SNNs) for image classification and segmentation tasks on images produced by conventional RGB cameras and event-based cameras (in collaboration with Prof. E. Chicca)
• evaluate their robustness to perturbations inexperienced in the training phase and against adversarial attacks.

2. PhD Project 2 (PP2). Daily supervisor: Prof. Michael Biehl; Co-supervisors: Dr. George Azzopardi and Prof. Kerstin Bunte

• study learning processes systematically in model situation by applying methods borrowed from statistical mechanics and the physics of disordered systems
• investigate the role of specific loss functions and hidden unit activations in layered neural networks with respect to their training dynamics and performance
• consider activation functions and training strategies which are inspired by realizable neuromorphic systems and favor sparse activity or connectivity.

The objective of each of these temporary positions is the production of a number of research articles in peer-reviewed scientific journals and conference proceedings, which together will form the basis of a thesis leading to a PhD degree (Dr) at the University of Groningen.

Qualifications

The successful candidate of PP1 should:

• have a keen interest in pursuing fundamental research in deep learning with a focus on brain-inspired high-energy efficient methodologies, with the aim to achieve sparsity, generalization beyond the distribution of the training set and robustness against adversarial attacks
• have a master’s degree or equivalent in computer science, artificial intelligence or another relevant field
• have proven experience with building deep learning architectures with Python for computer vision problems.

The successful candidate of PP2 should:

• display a deep interest in the fundamental understanding of neural networks and training processes in machine learning
• hold an MSc degree in (theoretical) physics or a related discipline with a strong background in statistical physics
• have experience with analytical and computational methods from the physics of disordered systems and dynamical systems, ideally in the context of machine learning
• be open to interaction and collaboration with material science oriented researchers within the CogniGron research center.

Candidates for both PP should have excellent verbal and written communication skills in English. They should possess good analytical skills and have a positive attitude towards collaborating with the PhD student on the complementary PP.

Conditions of employment

For each of the two PhD projects we offer, following the Collective Labour Agreement for Dutch Universities:

• a salary of € 2,541 gross per month in the first year, up to a maximum of € 3,247 gross per month in the fourth and final year for a full-time working week
• a holiday allowance of 8% gross annual income and an 8.3% year-end bonus
• a full-time position (1.0 FTE). The successful candidate will first be offered a temporary position of one year with the option of renewal for another three years. Prolongation of the contract is contingent on sufficient progress in the first year to indicate that a successful completion of the PhD thesis within the next three years is to be expected. A PhD training programme is part of the agreement and the successful candidate will be enrolled in the Graduate School of Science and Engineering.

Information

For information you can contact:

Dr George Azzopardi,   [email protected]

Prof Michael Biehl,   [email protected]

(please do not use the email addresses above for applications)

Additional information
Prof Michael Biehl
Dr George Azzopardi