Abstract
An important challenge in machine learning is performing with accuracy when few training samples are available from the target distribution. If a large number of training samples from a related distribution are available, transfer learning can be used to improve the performance. This paper investigates how to do transfer learning more effectively if the source and target distributions are related through a Sparse Mechanism Shift for the application of next-frame prediction. We create Sparse Mechanism Shift-TempoRal Intervened Sequences (SMS-TRIS), a benchmark to evaluate transfer learning for next-frame prediction derived from the TRIS datasets. We then propose to exploit the Sparse Mechanism Shift property of the distribution shift by disentangling the model parameters with regard to the true causal mechanisms underlying the data. We use the Causal Identifiability from TempoRal Intervened Sequences (CITRIS) model to achieve this disentanglement via causal representation learning. We show that encouraging disentanglement with the CITRIS extensions can improve performance, but their effectiveness varies depending on the dataset and backbone used. We find that it is effective only when encouraging disentanglement actually succeeds in increasing disentanglement. We also show that an alternative method designed for domain adaptation does not help, indicating the challenging nature of the SMS-TRIS benchmark.
Talk to us
Join us for a 30 min session where you can share your feedback and ask us any queries you have