Reproduce “Transferability Metrics for Selecting Source Model Ensembles” on OpenResearch

About this paper

Andrea Agostinelli, Jasper Uijlings, Thomas Mensink, Vittorio Ferrari

Google Research introduces novel transferability metrics for selecting ensembles of pre-trained models in transfer learning, which allow for efficient identification of model combinations without extensive fine-tuning. This approach significantly boosts performance in semantic segmentation, outperforming even single large models of equivalent capacity.

Key insights

• Transferability metrics can accurately predict the performance of source model ensembles, not just individual models, on target tasks.
• Combining diverse lightweight pre-trained models into an ensemble yields better performance than the single best model, and even a single larger model of equivalent parameter capacity.
• Diversity among ensemble members, in terms of source datasets, model architectures, and pre-training schemes, is critical for achieving synergistic performance gains.

Summary by alphaXiv.