MIRA / Publication

Publication

Foundation ModelsTheory in Dynamical SystemsRecurrent SystemsParallel Architecture

2026

MIRA: Mitigating Pole Geometry Bias in Recurrent Language Models

Róisín Luo1, Ihsan Ullah1,2, Christian Gagné3,4,5, Dong Li6, and Karyn Morrissey1

1 University of Galway · 2 Research Ireland – D-REAL · 3 Université Laval (IID) · 4 Canada–CIFAR AI Chair · 5 Mila – Québec AI Institute · 6 PASA Lab – University of California, Merced

  • preprint
  • Dynamical Systems
  • Signal Processing
  • Language Models
  • Recurrent Architectures
  • Parallel Computing
Read
Bib
@misc{luo2026mira,
  title        = {{MIRA}: Mitigating Pole Geometry Bias in Recurrent Language Models},
  author       = {Luo, R{\'o}is{\'i}n and Ullah, Ihsan and Gagn{\'e}, Christian and Li, Dong and Morrissey, Karyn},
  year         = {2026},
  eprint       = {2611.09412},
  archivePrefix = {arXiv},
  primaryClass = {cs.AI},
  url          = {https://arxiv.org/abs/2611.09412}
}

Lipschitz Continuity in Deep Learning: A Systematic Review of Theoretical Foundations, Estimation Methods, Regularization Approaches, and Certifiable Robustness

Róisín Luo1,2,3,4, James McDermott1,3,4, and Colm O’Riordan1,3,4

1 Research Ireland – Centre for Research Training in AI (CRT-AI) · 2 J.E. Cairnes School of Business & Economics · 3 School of Computer Science · 4 University of Galway, Ireland

  • TMLR2026
  • Theory
  • Lipschitz Continuity
  • Deep Learning
  • Training Optimisation
  • Certified Robustness
Read
Bib
@article{luo2026lipschitzreview,
  title        = {Lipschitz Continuity in Deep Learning: A Systematic Review of Theoretical Foundations, Estimation Methods, Regularization Approaches, and Certifiable Robustness},
  author       = {Luo, R{\'o}is{\'i}n and McDermott, James and O'Riordan, Colm},
  journal      = {Transactions on Machine Learning Research},
  year         = {2026},
  eprint       = {2607.16329},
  archivePrefix = {arXiv},
  primaryClass = {stat.ML},
  url          = {https://arxiv.org/abs/2607.16329}
}

A Stochastic–Geometric Theory of Scaling Laws in Grokking

Róisín Luo1,2, Christian Gagné3,4,5, Jonas Ngnawé3,5, Ihsan Ullah2,6, and Karyn Morrissey2

1 Research Ireland – Centre for Research Training in AI · 2 University of Galway · 3 Université Laval · 4 Canada–CIFAR AI Chair · 5 Mila – Québec AI Institute · 6 Visual Intelligence Lab

  • preprint
  • Theory
  • Training Optimisation
  • Dynamical Systems
  • Scaling Laws
  • Generalisation
Read
Bib
@misc{luo2026grokking,
  title        = {A Stochastic--Geometric Theory of Scaling Laws in Grokking},
  author       = {Luo, R{\'o}is{\'i}n and Gagn{\'e}, Christian and Ngnaw{\'e}, Jonas and Ullah, Ihsan and Morrissey, Karyn},
  year         = {2026},
  eprint       = {2606.30388},
  archivePrefix = {arXiv},
  primaryClass = {stat.ML},
  url          = {https://arxiv.org/abs/2606.30388}
}

2025

Higher-order singular-value derivatives of real rectangular matrices

Róisín Luo1,2, Colm O’Riordan1,2, and James McDermott1,2

1 University of Galway, Ireland · 2 Research Ireland – Centre for Research Training in Artificial Intelligence (CRT-AI), Ireland

  • JMAA2025
  • Theory
  • Operator Theory
  • Matrix Analysis
  • Spectral Theory
  • Perturbation Theory
Read
Bib
@article{luo2026singularvaluederivatives,
  title   = {Higher-order singular-value derivatives of real rectangular matrices},
  author  = {Luo, R{\'o}is{\'i}n and O'Riordan, Colm and McDermott, James},
  journal = {Journal of Mathematical Analysis and Applications},
  year    = {2026},
  volume  = {556},
  number  = {2},
  pages   = {130236},
  doi     = {10.1016/j.jmaa.2025.130236},
  url     = {https://www.sciencedirect.com/science/article/pii/S0022247X25010170}
}

Optimization-Induced Dynamics of Lipschitz Continuity in Neural Networks

Róisín Luo1,2, James McDermott1,2, Christian Gagné3,4,5, Qiang Sun6,7, and Colm O’Riordan1,2

1 University of Galway, Ireland · 2 Irish National Centre for Research Training in AI (CRT-AI) · 3 Université Laval, Canada · 4 Canada-CIFAR AI Chair · 5 Mila - Québec AI Institute · 6 University of Toronto, Canada · 7 MBZUAI, UAE

  • preprint
  • Theory
  • Training Optimisation
  • Dynamical Systems
  • Lipschitz Continuity
  • Deep Learning
Read
Bib
@misc{luo2025lipschitzdynamics,
  title        = {Optimization-Induced Dynamics of Lipschitz Continuity in Neural Networks},
  author       = {Luo, R{\'o}is{\'i}n and McDermott, James and Gagn{\'e}, Christian and Sun, Qiang and O'Riordan, Colm},
  year         = {2025},
  eprint       = {2506.18588},
  archivePrefix = {arXiv},
  primaryClass = {cs.LG},
  url          = {https://arxiv.org/abs/2506.18588}
}

2024

Reclaiming Residual Knowledge: A Novel Paradigm to Low-Bit Quantization

Roisin Luo1, Alexandru Drimbarean2, James McDermott1, and Colm O’Riordan1

1 University of Galway · 2 FotoNation

  • BMVC2024
  • Model Compression
  • Quantisation
  • Low-Rank Adaptation
  • Computer Vision
  • Neural Networks
Read
Bib
@inproceedings{Luo_2024_BMVC,
author    = {Roisin Luo and Alexandru Drimbarean and James McDermott and Colm O'Riordan},
title     = {Reclaiming Residual Knowledge: A Novel Paradigm to Low-Bit Quantization},
booktitle = {35th British Machine Vision Conference 2024, {BMVC} 2024, Glasgow, UK, November 25-28, 2024},
publisher = {BMVA},
year      = {2024},
url       = {https://papers.bmvc2024.org/0184.pdf}
}

Interpreting Global Perturbation Robustness of Image Models using Axiomatic Spectral Importance Decomposition

Róisín Luo1,2, James McDermott1,2, and Colm O’Riordan1,2

1 SFI Centre for Research Training in Artificial Intelligence · 2 School of Computer Science, University of Galway

  • TMLR2024
  • Interpretability
  • Robustness
  • Computer Vision
  • Signal Processing
  • Feature Attribution
Read
Bib
@article{luo2024spectralimportance,
  title   = {Interpreting Global Perturbation Robustness of Image Models using Axiomatic Spectral Importance Decomposition},
  author  = {Luo, R{\'o}is{\'i}n and McDermott, James and O'Riordan, Colm},
  journal = {Transactions on Machine Learning Research},
  year    = {2024},
  url     = {https://arxiv.org/abs/2408.01139}
}

2021

Optimizing Large-Scale Plasma Simulations on Persistent Memory-based Heterogeneous Memory with Effective Data Placement Across Memory Hierarchy

Jie Ren1, Roisin Luo1, Ivy Peng2,3, Kai Wu1, and Dong Li1

1 University of California, Merced · 2 Lawrence Livermore National Laboratory · 3 KTH Royal Institute of Technology, Sweden

  • SC2021
  • High-Performance Computing
  • Scientific Computing
  • Memory Systems
  • Data Placement
  • Performance Optimisation
Read
Bib
@inproceedings{ren2021plasmamemory,
  title     = {Optimizing Large-Scale Plasma Simulations on Persistent Memory-based Heterogeneous Memory with Effective Data Placement Across Memory Hierarchy},
  author    = {Ren, Jie and Luo, Roisin and Peng, Ivy and Wu, Kai and Li, Dong},
  booktitle = {Proceedings of the 35th ACM International Conference on Supercomputing},
  year      = {2021},
  pages     = {203--214},
  publisher = {ACM},
  doi       = {10.1145/3447818.3460356},
  url       = {https://doi.org/10.1145/3447818.3460356}
}

Sentinel: Efficient Tensor Migration and Allocation on Heterogeneous Memory Systems for Deep Learning

Jie Ren1, Roisin Luo1, Kai Wu1, Minjia Zhang2,3, Hyeran Jeon1, and Dong Li1

1 University of California, Merced · 2 Microsoft Research · 3 University of Illinois Urbana-Champaign

  • HPCA2021
  • Computer Architecture
  • Memory Systems
  • Deep Learning
  • Training Optimisation
  • Runtime Systems
Read
Bib
@inproceedings{ren2021sentinel,
  title     = {Sentinel: Efficient Tensor Migration and Allocation on Heterogeneous Memory Systems for Deep Learning},
  author    = {Ren, Jie and Luo, Roisin and Wu, Kai and Zhang, Minjia and Jeon, Hyeran and Li, Dong},
  booktitle = {2021 IEEE International Symposium on High-Performance Computer Architecture (HPCA)},
  year      = {2021},
  pages     = {598--611},
  publisher = {IEEE},
  doi       = {10.1109/HPCA51647.2021.00057},
  url       = {https://doi.org/10.1109/HPCA51647.2021.00057}
}

2020

Enabling Faster NGS Analysis on Optane-Based Heterogeneous Memory

Roisin Luo1, Luanzheng Guo1, Jie Ren1, Kai Wu1, and Dong Li1

1 University of California, Merced

  • SC2020 (poster)
  • High-Performance Computing
  • Bioinformatics
  • Memory Systems
  • Persistent Memory
  • I/O Optimisation
Read
Bib
@misc{luo2020ngsoptane,
  title        = {Enabling Faster {NGS} Analysis on Optane-Based Heterogeneous Memory},
  author       = {Luo, Roisin and Guo, Luanzheng and Ren, Jie and Wu, Kai and Li, Dong},
  year         = {2020},
  howpublished = {Research poster at {SC20}: The International Conference for High Performance Computing, Networking, Storage and Analysis},
  url          = {https://sc20.supercomputing.org/proceedings/tech_poster/poster_files/rpost119s2-file3.pdf}
}