Self-supervised TICA

[Kai-Zhu-2001]

Description

While DeepTICA attempts to directly optimize the entire TICA objective, this often results in unstable training dynamics. Alternatively, a self-supervised learning framework that first learns a representation (i.e., model output) and subsequently applies TICA to be more stable and can produce more reliable and physically meaningful collective variables.

The CV implementation can be found in mlcolvar/cvs/timelagged/selftica.py. The training procedure is quite similar to that of DeepTICA; however, the TICA computation is not performed during the forward pass of training—it is only applied during inference. The corresponding loss functions are located in mlcolvar/core/loss/regspectral_loss.py.

Todos

• Check everything
• Write documentation and a tutorial

Questions

• Since self-supervised learning only provides a representation, the forward pass does not include the actual TICA step—so the output cannot be directly used as a CV. In my code, I registered a buffer to store the TICA eigenvectors during training and then use them during inference. Could you check if there is a better implementation?
• Is it reasonable to name the method SelfTICA?
• Because the master branch of mlcolvar has not yet implemented GNN, I did not upload the GNN codes.

Status

• Ready to go