vii only limited supervision on transformations, and (3) various desirable properties expressed by existing disentanglement metrics are also achieved by LSBD representations. • Lastly, we investigate how well both LSBD-VAE and traditional VAE-based disentanglement models can perform OOD generalisation in a number of controlled settings. Results show that both model types struggle with generalisation in more challenging settings. However, we also observe that the LSBD-VAE encoder often still learns a meaningful mapping that reflects the underlying group structure. In other words, the encoder may generalise well to data with unseen factor combinations even if the decoder struggles to correctly reconstruct this data.
RkJQdWJsaXNoZXIy MjY0ODMw