Continuously Streaming Artificial Intelligence

The talk will address one of the most challenging adversities faced by artificial intelligence systems, represented by forgetting previously learnt information when the AI system is trained on a new dataset. This is also called catastrophic forgetting. The ways to address catastrophic forgetting, is also called continual/lifelong learning, this is a fundamental property of leaving beings for adapting to new environments through continuous learning. The problem considered consists of learning a series of tasks, usually defined as classification and image generation problems when training on sequences of different datasets. The talk will discuss some of the approaches to continual learning, including by using expanding mixtures of models and teacher-student architectures. Mechanisms used in continual learning such as the Generative replay mechanism will be presented. Results on successively learning sequences of image datasets, for classification as well as generation applications will be provided. In conclusion the results so far as well as the limitations of current models will be discussed.

Post Talk Link:  Click Here 

Passcode: 0PX$BCU%