Test hypothesis on toy problem - Githubissues

dgcnz / dl2

Code for "Effect of equivariance on training dynamics"

1 stars 0 forks source link

Test hypothesis on toy problem #27

Closed dgcnz closed 2 months ago

dgcnz commented 2 months ago

Tasks:

[x] Implement RotMNIST dataloader
[x] Implement CNN baseline
[x] Implement Steerable GCNN
[x] Plot hessian eigenspectra

Results:

Neither model has negative eigenvalues #14
GCNN has smaller eigenvalues than CNN (better, expected?)
CNN eigenvalues are too high #26

dgcnz commented 2 months ago

Work done by @MeneerTS @Nesta-gitU