CL ² R: Compatible Lifelong Learning Representations

In this article, we propose a method to partially mimic natural intelligence for the problem of lifelong learning representations that are compatible. We take perspective agent is interested in recognizing object instances an open dynamic universe way which any update its internal feature representation does not render features gallery unusable visual search. refer as Compatible Lifelong Learning Representations (CL 2 R), it considers compatible within paradigm. identify stationarity property required hold achieve compatibility and novel training procedure encourages local global on learned representation. Due stationarity, statistical properties do change over time, making them interoperable with previously features. Extensive experiments standard benchmark datasets show our CL R outperforms alternative baselines state-of-the-art methods. also provide metrics specifically evaluate under catastrophic forgetting various sequential tasks. Code available at https://github.com/NiccoBiondi/CompatibleLifelongRepresentation .