Potential extensions

[Hrovatin]

• Weight cycle consistency by inverse system proportions to improve integration of rare systems in multi-system settings
• Could we disable gradients in the cycle decoding and encoding so that only the encoder of real data gets updated via cycle loss?
  • Potential pros: Ensures that the real data encoding and not cycle data decoding/encoding is adjusted via cycle consistency loss
  • Potential cons: The encoder/decoder then could not learn how the mock covariates used only in cycle look like. Could instead use real covariates of the original input, but in that case, the decoder/encoder may not know how to deal with them in the new system as many covariate values are system-specific (e.g. within-system batches)