Breaking Free from the Stability-Plasticity Dilemma with Incremental Domain Inference on Sequential Data
Romain Mouret · viXra · 2021
We make the case for identifying the input domain prior to running downstream models and propose an architecture that opens the door to lifelong learning systems that forget at a decreasing rate as the tasks grow in complexity. Our model accurately identifies domains and is compatible with other continual learning algorithms, provided they benefit from knowing the current domain beforehand.