Claudia Clopath (Imperial College London)

Continual learning though consolidation – a neuroscience angle

Abstract:

I will review the different mechanisms the brain might use to mitigate catastrophic forgetting in the brain and present a couple of brain-inspired agents in a reinforcement learning set up.

Jeff Clune (OpenAI, University of Wyoming)

Learning to continually learn

Abstract:

A dominant trend in machine learning is that hand-designed pipelines are replaced by higher-performing learned pipelines once sufficient compute and data are available. I argue that trend will apply to machine learning itself, and thus that the fastest path to truly powerful AI is to create AI-generating algorithms (AI-GAs) that on their own *learn* to solve the hardest AI problems. This paradigm is an all-in bet on meta-learning. After introducing these ideas, the talk focuses on one example of this paradigm: Learning to Continually Learn. I describe a Neuromodulated Meta-Learning algorithm (ANML), which uses meta-learning to try to solve catastrophic forgetting, producing state-of-the-art results.

Alexei (Alyosha) Efros (UC Berkley)

Imagining a Post-Dataset Era

Abstract:

Large-scale datasets have been key to the progress in fields like computer vision during the 21st century. Yet, the over-reliance on datasets has brought new challenges, such as various dataset biases, fixation on a few standardized tasks, failure to generalize beyond the narrow training domain, etc. It might be time to move away from the standard training set / test set paradigm, and consider data as it presents itself to an agent in the real world -- via a continuous, non-repeating stream. In this talk, I will discuss some of the potential benefits, as well as the challenges, of learning in a post-dataset world, including some of our recent work in test-time training.

Christoph H. Lampert (IST Austria)

Learning Theory for Continual and Meta-Learning

Abstract:

In recent years we have seen an explosion of approaches that aim at transferring information between different learning tasks, in particular meta-learning and continual or lifelong learning. In my talk, I discuss ways to study these formally, using tools from learning theory that abstract away the specific details of implementation. In particular, I will discuss which assumptions one has to make on the tasks to be learned in order to guarantee a successful transfer of information.

Bing Liu (University of Illinois at Chicago)

Learning on the Job in the Open World

Abstract:

In existing machine learning (ML) applications, once a model is built it is deployed to perform its intended task. During the application, the model is fixed due to the closed-world assumption of the classic ML paradigm – everything seen in testing/application must have been seen in training. However, many real-life environments - such as those for chatbots and self-driving cars - are full of unknown, which are called the open environments/worlds. We humans can deal with such environments comfortably - detecting unknowns and learning them continuously in the interaction with other humans and the environment to adapt to the new environment and to become more and more knowledgeable. In fact, we humans never stop learning. After formal education, we continue to learn on the job or while working. AI systems should have the same on-the-job learning capability. It is impossible for them to rely solely on manually labeled data and offline training to deal with the dynamic open world. This talk discusses this problem and presents some initial work in the context of natural language processing.

Razvan Pascanu (DeepMind)

Continual Learning from an Optimization/Learning-dynamics perspective

Abstract:

Continual learning is usually described through a list of desiderata, however some of the "wants" on this list are in contradiction with each other, hence a solution to continual learning implies finding suitable trade-offs between the different objectives. Such trade-offs can be given by grounding ourselves into a particular domain or set of tasks. Alternatively, I believe, one can also rely on framing or looking at continual learning through different perspectives to gain this grounding. In this talk I'm looking at optimization and learning dynamics. From this perspective, continual learning can be seen as looking for a more suitable credit assignment mechanism for learning, one that does not rely on tug-of-war dynamics that result from gradient based optimization techniques. I exemplify in what sense this grounds us, and present a few recent projects I've been involved in that could be thought of as looking at continual learning from this perspective.