Unverified50% confidenceFactExact time
Gradient analysis found that adversarial optimization is highly concentrated on a small number of parameters (robustness-critical parameters, RCPs), with robustness encoded in a sparse subspace.
1
Sources
50%
Confidence
Long-term
Relevance
10/30/2026
First Seen
Sources
Related Entities
Related Claims
Unverified无梯度方法样本效率低，在高维参数空间中缺乏梯度指引意味着需要海量函数评估才能收敛73% similarUnverifiedIn the special case of a uniform design, this study obtains matching upper and lower bounds for pointwise CATE estimation at interior points, but the result is limited to the low-nuisance-smoothness regime.72% similarUnverifiedSAFT (Sparse Adversarial Fine-Tuning) identifies RCPs through adversarial gradient saliency, updates only these parameters, and freezes the rest.70% similarUnverifiedStandard Bayesian optimization acquisition functions (such as GP-UCB, EI, Thompson sampling) are almost all designed to minimize cumulative regret, and therefore naturally favor exploitation over exploration.70% similarUnverifiedWhen the parameter space is generalized from finite dimensions to an infinite-dimensional function space, the inner optimization problem of maximizing the acquisition function becomes neither exhaustively searchable nor readily amenable to direct gradient descent.69% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/1114009API
curl https://kongchang.com/api/v1/knowledge/claims/1114009MCP
get_claim(id=1114009)