Researchers have developed a novel online sketched Newton method that uses a generalized accelerated sketch-and-project solver (GAS) to approximate Newton directions, addressing the computational bottleneck of traditional second-order methods. This GAS solver incorporates Nesterov momentum for accelerated convergence and allows for flexible projection metrics to reduce computational costs. The study establishes asymptotic normality of the averaged sketched Newton iterates and characterizes their limiting covariance matrix, showing it converges more rapidly and is smaller than that of the last iterate produced by the accelerated method. Furthermore, a functional central limit theorem is proven, enabling an online inference procedure based on random scaling that bypasses explicit covariance estimation and achieves asymptotically valid results, as demonstrated by numerical experiments. AI
IMPACT Introduces a more efficient inference procedure for optimization methods, potentially impacting AI model training.
RANK_REASON The cluster contains an academic paper detailing a new inference method for Newton methods. [lever_c_demoted from research: ic=1 ai=1.0]
AI-generated summary · Google Gemini · from 1 sources. How we write summaries →
