Clarkson University PhD Candidate Zander Blasingame and Professor of Electrical and Computer Engineering Chen Liu just had their recent paper accepted at the Thirty-Eighth Annual Conference on Neural Information Processing Systems (NeurIPS 2024). The paper is titled “AdjointDEIS: Efficient Gradients for Diffusion Models.”
“AdjointDEIS is a novel strategy for efficiently calculating the gradients of diffusion models w.r.t. to the solution trajectories, model parameters, and conditional information. This can be used for a variety of applications and we illustrate it using an adversarial attack in the form of an image generation problem.” Blasingame said. “We present bespoke ODE/SDE solvers that use multi-step methods to efficiently solve the adjoint Probability Flow ODE or adjoint diffusion SDE.”
NeurIPS, founded in 1987, is one of the top three high impact conferences in machine learning and artificial intelligence in the world, along with ICLR and ICML. It is a globally renowned conference that attracts upwards of 15,000 attendees featuring leading researchers, engineers, and data scientists. This research is a key component of Blasingame’s Ph.D. dissertation and he is looking forward to presenting it in Vancouver this December.
More information about the paper can be found by clicking here.