PhD in Symmetries and their Breaking in Deep Learning

University of Amsterdam; Amsterdam; On-site; EUR 35,000 - 52,000; Full time

Job summary

The University of Amsterdam invites applications for a four-year PhD position focusing on symmetry breaking in deep learning and generative modeling. The project combines theory and empirical work, including algorithms and training of models on a HPC cluster.

You will teach labs and supervise bachelor and master theses; a MSc in AI/CS/Physics/Math is required, with strong math, Python, PyTorch or JAX, and English proficiency.

Qualifications

• MSc in artificial intelligence, computer science, physics, mathematics, or a related field.
• Strong mathematical skills and ability to build mathematical models of complex phenomena.
• Working knowledge of machine learning, including modern deep learning architectures such as transformers.
• Proficiency in Python or related languages.
• Familiarity with scientific libraries (e.g. NumPy/SciPy, PyTorch or JAX) is preferred.
• Interest in interdisciplinary research across AI/ML, CS, mathematics and physics.
• Fluent in English, both speaking and writing.
• Willingness to take on teaching tasks as a TA and supervise Bachelor’s and Master’s theses.

Responsibilities

• Conducting independent theoretical research, including development of novel algorithms and mathematical study of the dynamics of deep learning.
• Conducting independent empirical research, including coding, training and evaluating deep learning models on a high-performance computing cluster.
• Presenting research results at international conferences, workshops, and journals.
• Completing a PhD thesis within the appointed duration of four years.
• Assisting in teaching activities, such as teaching labs and tutorials or supervising bachelors and master students.

Skills

Mathematical modeling

Python

PyTorch

JAX

NumPy

Deep learning

Transformers

Machine learning

English fluency

Education

MSc in AI/CS/Physics/Math

Tools

PyTorch

JAX

NumPy

SciPy

Job description

Symmetry breaking in deep learning: a central problem in deep learning is the coherent propagation of information through many layers, and the integration of information across space and time to perform a computation. Nature solves this problem in various ways, which often include the spontaneous breaking of symmetries, a topic which lies at the heart of our understanding of the phases of matter in the natural world. Recent work has shown that some of the same principles can be applied to deep learning. Can we exploit these ideas to construct new physically-inspired neural network architectures for efficient computation? Generative models and non-reciprocal thermodynamics: despite their amazing empirical successes, the fundamental principles underlying the operation of generative models remains somewhat unclear. Is there a theory for how generative models such as diffusion models and autoregressive text generation generalize from their training set? Can simple dynamic principles constrain the creation of structure in such a generative model? Can we obtain insights into the process of generalization from fundamental principles in statistical physics and use these insights to build new kinds of generative models.

Conducting independent theoretical research, including development of novel algorithms and mathematical study of the dynamics of deep learning Conducting independent empirical research, including coding, training and evaluating deep learning models on a on a high-performance computing cluster. Presenting research results at international conferences, workshops, and journals. Completing a PhD thesis within the appointed duration of four years Assisting in teaching activities, such as teaching labs and tutorials or supervising bachelors and master students.

You have an MSc in artificial intelligence, computer science, physics, mathematics, or a related field. You have strong mathematical skills and are comfortable with building mathematical models of complex phenomena. You have a working knowledge of machine learning, including modern deep learning architectures such as transformers. You have sufficient programming skills in Python or related languages. Familiarity with scientific libraries (e.g. NumPy/SciPy, PyTorch or JAX) is preferred. You are interested in interdisciplinary research, crossing AI/Machine Learning, computer science, mathematics, and physics You are fluent in the English language, both in speaking and writing. You are willing to take a light education task, as teaching assistant and in supervising Bachelor’s and Master’s students in their thesis research work.

Nabil Iqbal, Associate Professor, [e-mail]