Felix Wiewel

Deep Learning - Continual Learning

Neural networks achieve state-of-the-art performance and in some cases even surpass humans on many machine learning tasks. Despite their success they lack some important properties compared to the way humans and animals learn. One of these is the ability to learn tasks in a sequential order where only training examples of the current tasks are available. Currently neural networks suffer from a drastic decrease in performance on previous tasks when they are trained on a new task. This phenomenon is know as "catastrophic forgetting". Research in the area of continual learning aims at preventing catastrophic forgetting in an efficient and salable way while at the same time enabling the transfer of knowledge obtained on previous into future tasks and vice versa. This ability of learning tasks sequentially and benefiting from other learned tasks is considered by many as an important step on the way to a general artificial intelligence.