Preconditioned-Stochastic-Gradient-Descent

Join the discord

Quick Guide

python3 psgd_cifar10.py --experiment cifar10 --optimizer PSGD_XMat

You can pick from the following CIFAR10 --experiment:

• Standard: cifar10
• Class Imballanced: imb
• NTK Attacked: attacked
• Noisy Label: noisy
• Blurred: blurred

You can control if you want to change dataloaders on the fly by setting a --stage2 dataset. For example:

python3 psgd_cifar10.py --experiment blurred --stage2 cifar10 --epoch_concept_switch 100  --optimizer PSGD_XMat --lr_scheduler exp

will train a ResNet-18 for 100 epochs and then switch to training with standard clean cifar10 data. Note lr_scheduler of exp is to be consistant with Critical Learning Period's paper and does not yield the best results. For best results use --lr_scheduler cos

and

python3 psgd_cifar10.py --experiment blurred --stage2 cifar10 --epoch_concept_switch 100  --optimizer SGD --num_runs 5

will train a ResNet-18 for 100 epochs and then switch to training with standard clean cifar10 data using SGD

For NTK Attacked dataset you need to download and set the path via the --data_root argument.

If you want to run the Noisy experiments that uses proir information run the psgd_cifar10_noisy_label.py code.

Experiment Observations

Noisy Label : Pure Memorization without Generalization

• PSGD gets test acc 78% avg over 5 nets with ~45% train acc over noisy labels.
• SGD gets test acc 23% avg over 5 nets with ~44% train acc over noisy labels.
  • 4/5 get 10% test acc at 200 epochs with 99.99% confidence in predictions
    • Pure Memorization -- Simply overfit the train set but with no generalization to the test set 10% accuracy on test.
    • With a bad teacher seems the best they can do is memorize; since they get 10% acc on test set with super high confidence
  • 1/5 gets 77% test acc at 200 epochs with 99.99% confidence in predictions
    • Lucky Initilization -- actually super smart and can learn/generalize even given a teacher thats wrong 54% of the time
  • both have ~44% acc on noisy labeled train set

Blurred: Clear indication of PSGD retaining neuro-plasticity vs SGD.

• Train for 100 epochs of blur and then for another 100 with standard:
  • PSGD recover test accuracy of 93.5% with cosine lr sched *a 2% decrease compared to no deficit a ~1% decrease for SGD
  • With exp decay lr sched and removing the deficit at 100 epochs:
    • PSGD got
    • While the reported numbers for SGD was abount 84%

Dataset Setup:

Download the Neural Tangent Generalization Attacks Dataset and put it in the datasets folder.

TODO

• Integrate Trace of FID
• Integrate entropy max margin and forgetting
• Add RL Experiments & Results
• add SimCLR Experiments & Results
• add ConvMix Experiments & Results
• add ViT Experiments & Results
• add NAS Experiments & Results