As neural networks are increasingly being applied to real-world applications,\nmechanisms to address distributional shift and sequential task learning without\nforgetting are critical. Methods incorporating network expansion have shown\npromise by naturally adding model capacity for learning new tasks while\nsimultaneously avoiding catastrophic forgetting. However, the growth in the\nnumber of additional parameters of many of these types of methods can be\ncomputationally expensive at larger scales, at times prohibitively so. Instead,\nwe propose a simple task-specific feature map transformation strategy for\ncontinual learning, which we call Efficient Feature Transformations (EFTs).\nThese EFTs provide powerful flexibility for learning new tasks, achieved with\nminimal parameters added to the base architecture. We further propose a feature\ndistance maximization strategy, which significantly improves task prediction in\nclass incremental settings, without needing expensive generative models. We\ndemonstrate the efficacy and efficiency of our method with an extensive set of\nexperiments in discriminative (CIFAR-100 and ImageNet-1K) and generative (LSUN,\nCUB-200, Cats) sequences of tasks. Even with low single-digit parameter growth\nrates, EFTs can outperform many other continual learning methods in a wide\nrange of settings.\n