Continual Backprop: Stochastic Gradient Descent with Persistent Randomness

arxiv:2108.06325

Original source · Paper JSON

Data, training, optimization · 2021-08-13

Cites (2) · Cited by (3)

Complete cleaned Markdown · Source Markdown including bibliography

Read in parts

1. Part 1
2. Part 2
3. Part 3

Contents

• Continual Backprop: Stochastic Gradient Descent with Persistent Randomness
• Abstract
• 1 Introduction
• 2 Non-stationary problems
• 2.1 Bit-Flipping problem
• 2.2 Permuted MNIST
• 2.3 Slippery Ant
• 3 Backprop loses the ability to adapt under extended tracking
• 4 Continual Backprop
• 5 Related works
• 6 Conclusion and limitations
• Appendix
• Appendix A More experiments with Backprop on the Bit-Flipping Problem
• Appendix B More experiments with CBP on the Bit-Flipping Problem
• Appendix C Ablation Study for the utility measure
• Appendix D Continual PPO
• Appendix E Compute Used
• Appendix F Affect of the speed of the distribution change on decaying plasticity
• Appendix G A deeper look in the networks: Feature Saturation and Vanishing Gradients
• Appendix H Affect of network size on decaying plasticity
• Acknowledgements.
• Instructions for reporting errors