Continual Adaptation of Visual Representations via Domain Randomization and Meta-learning

Most standard learning approaches lead to fragile models which are prone to\ndrift when sequentially trained on samples of a different nature - the\nwell-known "catastrophic forgetting" issue. In particular, when a model\nconsecutively learns from different visual domains, it tends to forget the past\ndomains in favor of the most recent ones. In this context, we show that one way\nto learn models that are inherently more robust against forgetting is domain\nrandomization - for vision tasks, randomizing the current domain's distribution\nwith heavy image manipulations. Building on this result, we devise a\nmeta-learning strategy where a regularizer explicitly penalizes any loss\nassociated with transferring the model from the current domain to different\n"auxiliary" meta-domains, while also easing adaptation to them. Such\nmeta-domains are also generated through randomized image manipulations. We\nempirically demonstrate in a variety of experiments - spanning from\nclassification to semantic segmentation - that our approach results in models\nthat are less prone to catastrophic forgetting when transferred to new domains.\n