|
Fred Roosta
- Professor
- School of Mathematics and Physics, University of Queensland, Australia
- Email:
fred.roosta@uq.edu.au
- Office:
Priestley Building (67) - Room 754
- Phone: +61 7 336 53259
|
News
2025
- February, 2025: A new paper on Hessian-aware scalings of gradeint descent: these scaled gradient directions inherit many of the interesting characteristics of Newton-type methods, such as natural scaling and the ability to use the unit step size locally, eliminating the need for a line search.
- February, 2025: A new paper on the uncertainty quantification of over-parameterized machine learning models using the empirical NTK and (stochastic) gradient descent.
- January, 2025: Our paper on the complexity guarantees for nonconvex Newton-MR under inexact Hessian information has been accepted for publication in the IMA Journal of Numerical Analysis.
2024
- November, 2024: I will be serving as Action Editor for the Transactions on Machine Learning Research (TMLR).
- November, 2024: I will be serving as an area chair for the 42nd International Conference on Machine Learning (ICML, 2025).
- October, 2024: Our paper on "Training-free Medical Image Inversion via Bi-level Guided Diffusion Models" has been accpected to IEEE/CVF Winter Conference on Applications of Computer Vision (WACV 2025)
- August, 2024: Vivak Patel (University of Wisconsin Madison) and I are organizing multiple sessions on the topic of Frontiers of Optimization for Machine Learning during the 2025 International Conference on Continuous Optimization (ICCOPT 2025), which will be held in Los Angeles from July 19 to July 24, 2025.
- August, 2024: I will be serving as an area chair for the 13th International Conference on Learning Representations (ICLR 2025).
- May, 2024: I will be serving as an area chair for the 38th Conference on Neural Information Processing Systems (NeurIPS, 2024).
- May, 2024: Two papers (this and this) have been accepted to the International Conference on Machine Learning (ICML 2024). Congratulations to Oscar Smee and Eslam Zaher on each having their debut papers accepted at ICML!
- April, 2024: New paper on guided diffusion models for medical imaging inverse problems.
- February, 2024: Our paper on an adaptive gradient descent appeared in the Transactions on Machine Learning Research (TMLR).
- January, 2024: Our new paper investigates conjugate direction methods, such as conjugate gradient (CG) and conjugate residual (CR), across a wide range of settings, including inconsistent ones. In particular, we demonstrate that the widely celebrated but more complex MINRES method is essentially equivalent to the much simpler CR method!
- January, 2024: In our new paper, we developed a new efficient algorithm for approximating leverage scores of a tall matrix, with applications to big time series data.
2023
- December, 2023: I will be serving as an area chair for the 41st International Conference on Machine Learning (ICML, 2024).
- November, 2023: A new paper on a variant of gradient descent in which the step-size is automatically tuned using a novel local first-order smoothness oracle, which generalizes the Lipschitz continuous gradients smoothness condition.
- November, 2023: A new paper on PAC-Bayesian bound for a general class of models, characterizing factors which influence generalization performance in the interpolating regime.
- October, 2023: I will be serving on the program committee of the 2025 SIAM Annual Meeting, to be held in Montreal, Canada, July 26 - August 2, 2025.
- October, 2023: In this paper, we show that for inconsistent linear systems, it is possible to obtain the pseudo-inverse solution in a variety of settings using a straightforward one-dimensional projection method.
- September, 2023: I will be serving as an area chair for the 12th International Conference on Learning Representations (ICLR 2024).
- August, 2023: Our recent paper on the iteration and operation complexity analysis of non-convex Newton-MR algorithm with inexact Hessian information is now available.
- July, 2023: In our new paper, we develop a (Bayesian) information criterion for interpolating overparameterized models that, among other things, captures the sharpness (flat minima) heuristic commonly used in deep learning.
- May, 2023: I will be serving on the program committee of the 2024 SIAM Conference on the Mathematics of Data Science (SIAM MDS), to be held October 21-25, 2024, in Atlanta, Georgia, USA.
- May, 2023: Our paper on uncertainty quantification of deep learning models applied to oncology has appeared in Nature's Scientific Reports.
- March, 2023: Our paper on double descent phenomenon in Gaussian processes has been accepted in ICML 2023.
- March, 2023: I will be serving as an area chair for the 37th Conference on Neural Information Processing Systems (NeurIPS) 2023.
2022
- December, 2022: I will be serving as an area chair for the 40th International Conference on Machine Learning (ICML) 2023.
- October, 2022: New paper on double descent phenomenon in Gaussian processes.
- September, 2022: Congratulations to my outgoing PhD student, Yang Liu, for landing a postdoc position at the Oxford Mathematical Institute as part of the Centre for Intelligent Multidimensional Data Analysis (CIMDA).
- August, 2022: Our paper on the complexity analysis of non-convex variants of Newton-MR is now available.
- July, 2022: Our paper on a variant of the Newton-CG method with inexact gradient and Hessian information has been accepted to IMA Journal of Numerical Analysis (IMAJNA).
- July, 2022: Our paper on a variant of the Newton's method with MINRES as sub-problem solver for invex optimization has been accepted to EURO Journal on Computational Optimization (EJCO).
- June, 2022: Our paper on negative curvature detection and monotonicity properties of MINRES has been accepted to SIAM Journal on Optimization (SIOPT).
- April, 2022: I will be serving on the organizing committee of the SIAM Conference on Optimization (OP23), to be held May 30-June 3, 2023, in Seattle, Washington, USA.
- March, 2022: I will be serving as an area chair for the 36th Conference on Neural Information Processing Systems (NeurIPS) 2022.
2021
- December, 2021: I will be serving as an area chair for the 39th International Conference on Machine Learning (ICML) 2022.
- November, 2021: Our
new
paper on a novel invexifying regualrization framework for non-linear least squares problems is available.
- September, 2021: Our
new
paper on complexity analysis of variants of Newton-CG algorithm with inexact gradient and Hessian for non-convex optimization is available.
- January, 2021: With Albert S. Berahas (University of Michigan), we will have a two part mini-symposium on "Beyond First Order Methods in Machine Learning" during the EUROPT 2021, the 18th international workshop on continuous optimization.
2020
- July, 2020: Our paper on
Newton-ADMM, a distributed GPU-accelerated second-order
optimization method, has been accepted in the Proceedings of
the ACM/IEEE Supercomputing Conference (SC20) - 18%
acceptance rate.
-
June, 2020:
Our
paper on central limit theorems and concentration
inequalities for out-of-sample extensions of the adjacency and
Laplacian spectral embeddings has been accepted to the Journal
of Machine Learning Research (JMLR), with minor revision.
- June, 2020: Our
paper on distributed non-convex Newton-type optimization
methods has been accepted in
ICML, 2020! Congratulations to my PhD student
Rixon
Crane for his second paper in an A* conference!
- February, 2020: Our
recent
paper on stochastic normalizing flows, an extension of
continuous normalizing flows for maximum likelihood estimation
and variational inference, using stochastic differential
equations is available.
- February, 2020: Our
new
paper on studying Gaussian processes arising from
infinitely wide neural networks with ELU and GELU activations
as well as analysing the fixed-point dynamics of iterated
kernels is available.
- January, 2020: Our
paper
on the theory and application of the reproducing Stein kernel
for a posteriori correction of Monte-Carlo sampling algorithms
is now available on arXiv.
- January, 2020: Our
paper
on the theoretical and practical properties of Monte-Carlo
sampling algorithms by implicit discretizations of Langevin
SDE has been accepted to the Journal of Machine Learning
Research (JMLR), with minor revision.
2019
- December, 2019: Our
paper
on an empirical study of second-order optimization methods in
deep learning has been accepted to the SIAM International
Conference on Data Mining (SDM20) -
19%
acceptance rate.
-
December, 2019:
Our new
paper,
which extends the existing results on Gaussian process
iterpretation of multi-layer perceptrons to richer families of
priors, is available.
-
November, 2019:
Our new
paper on the application of RandNLA and leverage score
sampling within the context of big-data time series is now
available.
- November, 2019:
I will be serving as a senior program committee member for the
29th International Joint Conference on Artificial Intelligence
(IJCAI 2020).
-
October, 2019:
Our paper on central limit theorems and concentration
inequalities for out-of-sample extensions of the adjacency and
Laplacian spectral embeddings is now
available.
-
September, 2019:
Our paper on stability analysis of Newton-MR with respect to
Hessian approximations is now
available.
- September, 2019:
DINGO Does Vancouver! Our paper on
distributed Newton-type methods for optimization of
invex objectives has been accepted to NeurIPS 2019!
-
June, 2019: Congratulations to my PhD
student, Rixon Crane, for being awarded the Best Poster Award
at
AMSI Optimise this year.
-
June, 2019: I was selected as a top 5%
reviewer for ICML 2019.
- May, 2019: After 17 months and 21 days, our
paper on non-convex Newton-type methods with inexact
Hessian information has been accepted to Mathematical
Programming.
-
May, 2019: Our
paper
on the invariance kernels of multilayer perceptrons during
training has been accepted to the International Joint
Conference on Artificial Intelligence (IJCAI
2019) -
18%
acceptance rate.
-
April, 2019: Our
paper
on the theoretical and practical properties of Monte-Carlo
sampling algorithms by implicit discretizations of Langevin
SDE is now available on arXiv.
|