EEC_CCL Versions

[WesleyHsieh0806]

V2
Observation:
CCL helps the model learn faster at the first beginning, but it hinders the decrease of CE loss at last few epochs

1. Fix the lambda

Not much difference: 0.776

[WesleyHsieh0806]

V3

Stop computing MSE between the CC matrix of z and recover_z. Instead, force the CC matrix of z to approach to identity matrix

Not bad reach 0.783

[WesleyHsieh0806]

V4
for the CC matrix between z and recover_z to identity matrix
no distinct improvement
Roughly:0.776

[WesleyHsieh0806]

V5
MSe for the CC matrix between z and recover_z to identity matrix + force the CC matrix of z to approach to identity matrix (v3+v4)
Can reach 0.783