Signature



The condition-number tax: same matrix, same condition number, three iterations or three thousand
Aug 2026 We published a new blog post on our quantum research page (QuanTAS), about the symbol that every quantum linear solver quotes its cost in — the condition number \(\kappa\) — and what it is actually measuring.
The Sample-Complexity Theory Behind Randomized Max-3-Cut Rounding
Jul 2026 We published a new blog post on our explore-quantum page, working out the theory behind the randomized sampling algorithm used elsewhere in our low-rank Max-K-Cut series.
Which layer runs the program? What decides where a transformer computes
Jul 2026 We put up a new blog post on our AI-OWLS page, riding a question that is suddenly everywhere.
Can we train a computer? Two ways to point at a memory
Jul 2026 We put up a new blog post on our AI-OWLS page — the first of a short series of small, ground-truthed observations about how trained networks actually compute.
When trainability outlasts expressibility: training quantum eigensolvers with far fewer gates
Jul 2026 We published a new blog post on our quantum research page (QuanTAS), on a small structural trick that makes variational quantum eigensolvers easier to train — and a theorem that explains why it keeps helping long after it “should” have stopped.
Which valley, and how deep: training atomic relaxation at a fraction of the memory
Jul 2026 We published a new blog post on our AI-OWLS page, on training machine-learned interatomic potentials to relax atomic structures — the expensive inner loop of materials discovery — without the usual accuracy-versus-memory trade-off.
Two ways to slim a model: remember vs recompute
Jun 2026 We published a new blog post on our AI-OWLS page, on how to shrink a neural network without making it dumber — and why the right way to prune depends on what a given part of the model is actually doing.
One rank at a time: cascading error dynamics in sequential learning
May 2026 We published a new blog post on our AI-OWLS page, summarising our recent arXiv preprint — accepted to TMLR — with Mahtab Alizadeh Vandchali and Fangshuo (Jasper) Liao on error propagation in sequential rank-one learning.
One model, two roles: emergent specialization in a shared recurrent Transformer
May 2026 We published a new blog post on our AI-OWLS page, summarising our recent arXiv preprint with Jucheng Shen and Barbara Su on whether a shared-weight recurrent Transformer can develop multiple internal roles on its own.
From PCA to LoRA: why fine-tuning could have been parallel all along
May 2026 We published a new blog post on our AI-OWLS page, summarising our recent arXiv preprint with Barbara Su and Fangshuo (Jasper) Liao on parallel rank-one deflation for low-rank fine-tuning.
GHOST: pruning Mamba2 by what each channel actually does
May 2026 We published a new blog post on our AI-OWLS page, summarising our ICML 2026 paper with Michael Menezes on state pruning for Mamba2 selective state-space models.
How a little Gaussian dust changes how a network learns
May 2026 We published a new blog post on our AI-OWLS page, summarising our recent paper with Afroditi Kolomvaki, Fangshuo (Jasper) Liao, Evan Dramko, and Ziyun Guang on training two-layer ReLU networks under multiplicative Gaussian input noise.
Why Stochastic Gradient Descent Stops Just Short of the Edge
Apr 2026 We published a new blog post on our AI-OWLS page, summarising our recent paper with Fangshuo (Jasper) Liao and Afroditi Kolomvaki on the Edge of Stochastic Stability.
Provable Acceleration of Nesterov's Momentum for Deep ReLU Networks
Apr 2026 We published a new blog post on our AI-OWLS page, summarising our ALT 2024 paper with Fangshuo (Jasper) Liao on provable accelerated convergence of Nesterov’s momentum for deep ReLU neural networks.
Spectral vs. Combinatorial for Max-3-Cut
Apr 2026 We published a new blog post on our explore-quantum page, comparing spectral and combinatorial construction heuristics for Max-3-Cut.
Randomized Rank-2 for Max-3-Cut — When Two Eigenvectors Beat One
Apr 2026 We published a new blog post on our explore-quantum page, introducing a 3-phase spectral pipeline for Max-3-Cut that outperforms simulated annealing on half of the graph families tested.
Rank-1 as a Building Block for Million-Node Max-3-Cut
Apr 2026 We published a new blog post on our explore-quantum page, describing how three algorithmic insights make rank-1 spectral methods practical at extreme scale for Max-3-Cut.
What Can 15 Obsolete GPUs Do for Combinatorial Optimization?
Mar 2026 We published a follow-up blog post on our explore-quantum page, describing our GPU-parallel implementation of the low-rank Max-K-Cut solver.
Exploiting Low-Rank Structure in Max-K-Cut Problems
Mar 2026 We published a new blog post on our explore-quantum page, describing our work on exploiting low-rank structure in Max-K-Cut problems.
Major update
May 2024 Several updates happened in the past months: 1 paper was accepted at ICLR 2024 on adaptive step size for federated learning; 1 paper was accepted at ICML 2024 on the first work on how error propagates in deflation PCA; 1 paper was accepted at ISIT 2024 on a fast algorithm for FixMatch; and 1 paper was accepted at ICRA 2024. This constitutes the update of OptimaLab for the past 5 months. More details are in the Publications webpage or Google Scholar.
Major update
Jan 2024 Several updates happened in the past months: 1 paper was accepted at ICRC with the best poster award (Congrats, David Quiroga); 1 paper was accepted at ALT 2024; 1 paper was accepted at CPAL 2024; 1 paper accepted at NeurIPS 2023; 1 paper accepted at ICCV 2023; 1 paper accepted at ACML 2023; 1 journal accepted at the International Union of Crystallography Journal; and 1 journal paper accepted at the Machine Learning Journal (Springer). This constitutes the update of OptimaLab for the past 5 months. More details are in the Publications webpage or Google Scholar.
Major update
Apr 2023 OptimaLab has received an Amazon Research Award (ARA) and a Microsoft Research Award! Thank you sponsors!
Major update
Apr 2023 3 papers accepted at AISTATS 2023; 1 paper accepted at ICASSP; 1 paper accepted at ICRA; 2 journal publications on GIST and fast quantum tomography: This constitutes the update of OptimaLab the past 5 months. More details in the Publications webpage or Google Scholar.
Journal on neural network training accepted at Transactions on Machine Learning Research (TMLR)
Aug 2022 Fangshuo (Jasper) Liao has led the effort to prove why and when one can achieve this by iteratively creating, training, and combining randomly selected subnetworks in deep learning. The journal is accepted at the Transactions on Machine Learning Research (TMLR) - details below.
Two papers accepted at UAI 2022
More...