Darwinian Model Upgrades: Model Evolving with Selective Compatibility

The traditional model upgrading paradigm for retrieval requires recomputing all gallery embeddings before deploying the new (dubbed as "backfilling"), which is quite expensive and time-consuming considering billions of instances in industrial applications. BCT presents first step towards backward-compatible upgrades to get rid backfilling. It workable but leaves a dilemma between feature discriminativeness new-to-old compatibility due undifferentiated constraints. In this work, we propose Darwinian Model Upgrades (DMU), disentangle inheritance variation evolving with selective backward forward adaptation, respectively. old-to-new heritable knowledge measured by old discriminativeness, features, especially those poor quality, are evolved lightweight manner become more adaptive latent space. We demonstrate superiority DMU through comprehensive experiments on large-scale landmark face recognition benchmarks. effectively alleviates new-to-new degradation at same time improving compatibility, rendering proper systems.Code: https://github.com/TencentARC/OpenCompatible.