Under the assumption that current task data is a nonlinear transformation of prior task data, the paper proves informative statistical recovery bounds for experience replay with data-dependent regularization and weights in continual learning.
[Yes] (b) Complete proofs of all theoretical results
9 Pith papers cite this work. Polarity classification is still indexing.
9
Pith papers citing it
citation-role summary
background 1
citation-polarity summary
years
2026 9verdicts
UNVERDICTED 9roles
background 1polarities
background 1representative citing papers
A loss-driven Bayesian active learning framework derives unique acquisition objectives from arbitrary losses, with analytic solutions available when the loss is a weighted Bregman divergence.
UniPROT reformulates uniform prototype selection as a partial optimal transport problem that yields a submodular objective admitting a greedy algorithm with (1-1/e) approximation guarantee.
HiSS sampling uses logistic bridging in a Metropolis-within-Gibbs framework to enable transitions between disconnected modes while preserving the target discrete distribution.
New conditional independence assumptions enable mixture proportion estimation and kernel tests for conditional independence without relying on irreducibility.
Defining the Rashomon set for dimension reduction enables interpretable, robust visualizations by aligning embeddings with known structure and extracting consistent local relationships across multiple good embeddings.
A generalization bound based on a new feature-label distortion concept guides optimization of feature alignment versus target fitting in cross-modal adaptation and yields better empirical performance.
AdaScale-TuRBO scales Gaussian process lengthscales with problem dimension D and trust region side length L to preserve kernel geometry and improve performance over standard TuRBO in high-dimensional settings.
DIVERSED relaxes the verification step in speculative decoding with a dynamic ensemble verifier to raise token acceptance rates and speed up inference while keeping output quality intact.
citing papers explorer
-
The Rashomon Effect for Visualizing High-Dimensional Data
Defining the Rashomon set for dimension reduction enables interpretable, robust visualizations by aligning embeddings with known structure and extracting consistent local relationships across multiple good embeddings.