Meta Learning

Transformers excel at in-context learning by leveraging minimax-optimal nonparametric learning, achieving near-optimal risk with sufficient pretraining data diversity.

This research enhances meta-learning robustness by theoretically grounding and practically improving tail-risk minimization, achieving improved fast adaptation in the task space.

SPARKLE: A single-loop primal-dual framework unifies decentralized bilevel optimization, enabling flexible heterogeneity-correction and mixed update strategies for improved convergence.

This paper introduces uniform meta-stability for meta-learning, providing tighter generalization bounds for convex and weakly-convex problems, addressing computational limitations of existing algorith…

Meta-learning solves hybrid deep generative model unidentifiability!

New computational method infers complex brain learning rules from experimental data, revealing active forgetting in reward learning.

MetaNCoV: Learn data-driven priors via non-injective change of variables for enhanced few-shot learning.

FG²U: a novel memory-efficient algorithm for unbiased stochastic approximation of meta-gradients in large-scale bi-level optimization, showing superior performance across diverse tasks.

New algorithm efficiently solves linear regression with heterogeneous data batches, handling diverse input distributions and achieving high accuracy with fewer samples.

Surrogate PAC-Bayes Learning (SuPAC) efficiently optimizes generalization bounds by iteratively optimizing surrogate training objectives, enabling faster and more scalable learning for complex models.

Softmax attention in transformers adapts its attention window to function Lipschitzness and noise, enabling efficient in-context learning.

Two novel first-order algorithms, FOSL and MemCS, efficiently solve multi-block minimax bilevel optimization problems, significantly improving performance in deep AUC maximization and robust meta-lear…

FasMe: a novel meta-learning approach delivers fast and sample-efficient precision matrix estimation, surpassing existing methods in accuracy and speed for small sample datasets.

Nash bargaining resolves hypergradient conflicts in fairness-aware meta-learning, boosting model performance and fairness.

AI discovers robust, biologically-plausible plasticity rules that self-organize and maintain neural circuits’ sequential activity, even with synaptic turnover.

GEPS enhances parametric PDE solver generalization by using adaptive conditioning, achieving superior performance with limited data.

AccBO: A new accelerated algorithm achieves O(ε⁻³) oracle complexity for stochastic bilevel optimization with unbounded smoothness, significantly improving upon existing O(ε⁻⁴) methods.

EBiL-HaDS, a novel adaptive domain scheduler, significantly boosts open-set domain generalization by strategically sequencing training domains based on model reliability assessments.

MetaDebias tackles hidden confounding in recommender systems using heterogeneous observational data, achieving state-of-the-art performance without expensive RCT data.

Metalearning a neural circuit mimics nonparametric Bayesian inference, enabling fast, accurate, open-set classification.