arXiv cs.AI / cs.LG / cs.CL·4d agoA Spectral Theory of Grokking: Weight Decay induces Feature Learning#grokking#feature-learning#ntkAI research
arXiv cs.AI / cs.LG / cs.CL·9d agoSharp Reconstruction Bounds for Autoencoders Using the Same Forward Map#autoencoders#deep-learning-theory#diffeomorphismsAI research1
arXiv cs.AI / cs.LG / cs.CL·9d agoSpecial Lagrangian cones in Deep Learning#calibrated-geometry#deep-learning-theory#differential-geometryAI research1
arXiv cs.AI / cs.LG / cs.CL·15d agoBenign Loss Landscapes Can Coexist with Worst-Case Hardness#computational-hardness#deep-learning-theory#learning-theoryAI research
arXiv cs.AI / cs.LG / cs.CL·18d agoWhen Does Scale-Invariant Optimization Become Unstable? An Exact Schedule Law with Weight Decay#deep-learning-theory#learning-rate-schedules#normalizationAI research
arXiv cs.AI / cs.LG / cs.CL·18d agoCurriculum Learning as Transport: Understanding Curricula with Wasserstein Geodesics#curriculum-learning#data-ordering#deep-learning-theoryAI research1