Aurelien Lucchi

Quantum extreme learning machines can be designed to achieve stable interpolation thresholds, even with correlated features, by leveraging finite measurement shot noise as a regularizer.

Static problem pools cause LLM reasoning post-training to stall as models saturate the dataset, but dynamically synthesizing frontier tasks via regret-guided procedural generation keeps learning signals saturated for sustained gains.

Local SGD's efficiency can be rigorously explained through bounded second-order heterogeneity, revealing nearly tight convergence rates for general convex objectives.

Forget tedious hyperparameter tuning: adaptive optimizers like DP-SignSGD and DP-Adam maintain performance across privacy levels, unlike DP-SGD whose learning rate plummets with increased privacy.

AdamW's decoupled weight decay prevents Neural Collapse, challenging the assumption that this phenomenon is universal across optimization methods.