AI Paper Atlas

Research map / Optimization / Optimizers

Gradient Flossing: Improving Gradient Descent through Dynamic Control of Jacobians

Rainer Engelken

NeurIPS 2023 · Optimization › Optimizers · approach: Constrained and non-backprop training · 2 citations (OpenAlex)

Read the paper

AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Measured effects reported in the paper

What was changedWhereMeasuredEffect
add gradient flossing before trainingVanilla RNN · Delayed copy taskmax reliably trainable delay: 20 -> 40Improved
add gradient flossing during training in addition to before (ablation)Vanilla RNN · Delayed temporal XOR d=70test accuracy: chance level -> >80%Improved
add preflossing only (no flossing during training) (ablation)Vanilla RNN · Delayed temporal XORtest accuracy: chance at d>=40 -> chance at d=70Mixed
flossing during training on 3-input XOR (ablation)Vanilla RNN · 3-input XOR taskmax solvable delay: preflossing fails for d>60Improved
continue gradient flossing throughout all training epochs (ablation)Vanilla RNN · temporal XORtest accuracyHurt
regularize spectral radius or one-step Jacobians instead of Lyapunov exponents (ablation)Vanilla RNN · delayed taskstrainabilityNo clear effect
increase number of flossed Lyapunov exponents k (ablation)Vanilla RNN · binary delayed XORtest accuracyImproved
flossing all N exponents before training only (ablation)Vanilla RNN · hardest XOR taskstest accuracyNo clear effect
add gradient flossingVanilla RNN · long-term Jacobiancondition numberImproved

Closest papers by meaning

Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.

Open scime draft

Get the next scime Dispatch and a monthly field update

Free. We send a confirmation link first; every email has a one-click unsubscribe. Privacy