This repository contains an implementation of the article Lottery Ticket Hypothesis
And an application of this hypothesis to reinforcement learning
- Supervised
- Implement iterative magnitude pruning (IMP)
- Test using toy net and dataset CIFAR10
- Test using VGG19 net and dataset CIFAR10
- Make it fast
- Reinforcement learning
- Implement DQN
- Test on classic gym environments (CartPole, LunarLander)
- Try IMP (layerwise/global) with DQN on classic problems
- Add IMP with reinitialization to some epoch after training
- Add early stop criterions
- Add Rescaling weight after reinit
- Add Global/Layerwise/ERK pruners
- Analyze the specifics of applying Lottery ticket to DQN (e.g. target function updates)
- Dynamic epochs
- DDPG? Dueling networks? Different RL architecures...
- Atari games?
- Compare with other articles
- Clean up this list
- Rewinding technique and stability analysis:
Stabilizing the Lottery Ticket Hypothesis, Linear mode connectivity and the lottery ticket hypothesis - Application of hypothesis to reinforcement learning:
Lottery tickets in RL and NLP - Early-bird lottery tickets:
Drawing early-bird tickets - Pruning in RL:
Policy Pruning and Shrinking for Deep Reinforcement Learning - Reliability in RL:
Measuring the Reliability of Reinforcement Learning Algorithms - CURL:
Contrastive Unsupervised Representations for Reinforcement Learning
Optimal Brain Surgeon--second derivatives
Learning both Weights and Connections--prune + tune
Dynamic Network Surgery--parameter importance + grow pruned?
Layerwise Optimal Brain Surgeon--layerwise second derivatives
Grow and Prune Tool--??
Adaptive sparse connectivity -- ?? TODO
Overparametrized networks provably optimized--gradient descent on overparametrized networks
Rethinking the Value of Network Pruning--structured with random reinit
Transformed l1 regularisation for learning sparse DNNs -- something about l1 reg
Revisiting l1 regularisation for connection pruning -- something about l1 reg
Deconstructing Lottery Tickets--lottery ticket signs + supermasks
Sparse Networks from Scratch--sparse momentum
Making All Tickets Winners -- ?? RIGL
On Iterative Neural Network Pruning--pruning methods summary
Proving the Lottery Ticket--??
Improving Reliability of Lottery Tickets--??
Pruning untrained neural networks -- ??