Generative Models in the Fourier DomainDescriptionThe Data Science Institute (DSI) at the University of Technology Sydney (UTS) is a world leader in data science and ethical artificial intelligence with a strong focus on industry engagement.
The DSI boasts over 200 PhD students and 100 academic staff, including award-winning researchers in AI.
In collaboration with Leonardo.Ai, we are advertising an industry PhD program that will tackle research questions at the bleeding edge of GenAI research, aiming to drive foundational research in the field through traditional publication outputs and enhance the Leonardo.Ai platform.
Project Aims The overarching aim of this PhD is to explore and leverage the frequency domain (and other under-explored representations of pixel space) in training generative AI models across modalities (e.g., image, video).
The aim is to mode-reverse the diffusion process by using an autoregressive transformer (or similar techniques) to iteratively predict the frequency coefficients and phase values of the frequencies of an image, from low to high.
In model training and inference, we observe low-frequency components being initiated in the early generation steps, indicating that the denoising/attention process favors a starting point of broad-scale structure.
Effectively, the Fourier domain is an underexplored space in training generative models, and we want to exploit these techniques to produce state-of-the-art capabilities.
PhD research students will have access to leading-edge compute facilities (e.g., H100s), a team of world-class research leaders across both UTS, Leonardo, and Canva, cutting-edge datasets to conduct research, and mentoring and guidance to complete a best-in-class PhD research program.
Eligibility We are looking for the best of the best!
You will be adequately qualified to enter the UTS PhD scheme with undergraduate/postgraduate qualifications in Computer Science, Data Science, AI, Physics, Mathematics, or equivalent.
Candidate Profile A strong background in programming, mathematics, statistics, physics, or signal processing.Knowledge and interest in Diffusion, GANs, LLMs, and generative AI, as demonstrated by independent project work in GitHub.Proficiency in modern deep learning frameworks such as Pytorch or Jax.Ideally, we are looking for candidates with previous experience in top-tier tech companies.
We, however, welcome applications from underrepresented groups and diverse backgrounds.Proficiency in English (both verbal and written).Location This PhD will be hosted at the University of Technology Sydney, and the student will spend significant time at the Leonardo.Ai headquarters, located on the waterfront in North Sydney (20 mins train ride from UTS).
Full Domestic Scholarship: Open to Australian citizens who will receive a full scholarship of $37k per year stipend + $10k per year top-up for 3.5 years.
Full International Scholarship: Open to all international applicants with a $37k per year stipend, $45k per year international fees, and $30k per year top-up for 3.5 years.
Student visas will be provided by UTS.
#J-18808-Ljbffr