from RLCT repository
- estimate_softmaxDNN.py
compare exact RCLT(Learning Coeciffient) of linear NN and nonlinear NN (softmax/Relu)
-
llt_rlct_estimator_for_softmax_DNN.py
RLCT of small nonlinear coefficient to compare linear matrix decomposition network(exact solutoin).
-
llc_rlct_estimator_for_softmax_networks_via_power_posteriors_sgld.py
outor interpolation of SGLD LLC extimation with NUTS/HMC SGLD extimaiton with pyto.
-
llc_training_trajectory.py
LLT selection obserbation during SGD update, small LLC saddlepoint is selected or not.
python estimate_softmaxDNN.py
- The Local Learning Coefficient: A Singularity-Aware Complexity Measure
- Differentiation and Specialization of Attention Heads via the Refined Local Learning Coefficient
- Quantifying degeneracy in singular models via the learning coefficient
- Estimating the Local Learning Coefficient
- Singular Learning Theory with Daniel Murfet
- ReLUネットワークにおける局所学習係数推定手法のモデル選択への応用
- Stochastic natural gradient descent draws posterior samples in function space
- Stochastic Gradient Descent as Approximate Bayesian Inference
- Saddle-to-Saddle Dynamics Explains A Simplicity Bias Across Neural Network Architectures
- https://github.com/suswei/RLCT
- https://github.com/edmundlth/scalable_learning_coefficient_with_sgld/tree/v1.0