Continual learning Workshop
This is the webpage of the Continual Learning Workshop organized at Neural Information Processing Systems (NeurIPS) 2018.
Continual learning (CL) is the ability to learn continually from a stream of experiential data, building on what was learnt previously, while being able to reapply, adapt and generalize it to new situations. CL is a fundamental step towards artificial intelligence, as it allows the learning agent to continually extend its abilities and adapt them to a continuously changing environment, a hallmark of natural intelligence. It also has implications for supervised or unsupervised learning. For example, if a dataset is not randomly shuffled, or the input distribution shifts over time, a learned model might overfit to the most recently seen data, forgetting the rest -- a phenomenon referred to as catastrophic forgetting, which is a core issue CL systems aim to address.
Continual learning is characterized in practice by a series of desiderata. A non-complete list of which includes:
- Online learning -- learning occurs at every moment, with no fixed tasks or data sets and no clear boundaries between tasks;
- Presence of transfer (forward/backward) -- the learning agent should be able to transfer and adapt what it learned from previous experience, data, or tasks to new situations, as well as make use of more recent experience to improve performance on capabilities learned earlier;
- Resistance to catastrophic forgetting -- new learning should not destroy performance on previously seen data;
- Bounded system size -- the agent’s learning capacity should be fixed, forcing the system to use its resources intelligently, gracefully forgetting what it has learned so as to minimize potential loss of future reward;
- No direct access to previous experience -- while the model can remember a limited amount of experience, a continual learning algorithm cannot assume direct access to all of its past experience or the ability to rewind the environment (i.e., t=0 exactly once).
In the first (2016) meeting of this workshop, the focus was on defining a complete list of desiderata of what a continual learning (CL) enabled system should be able to do. The focus of the 2018 workshop will be on:
- how to evaluate CL methods; and
- how CL compares with related ideas (e.g., life-long learning, never-ending learning, transfer learning, meta-learning) and how advances in these areas could be useful for continual learning.
In particular, different desiderata of continual learning seem to be in opposition (e.g., fixed model capacity vs non-catastrophic forgetting vs the ability to generalize and adapt to new situations), which also raises the question of what a successful continual learning system should be able to do. What are the right trade-offs between these different opposing forces? How do we compare existing algorithms in the face of conflicting objectives? ? What metrics are most useful to report? In some cases, trade-offs will be tightly defined by the way we choose to test the algorithms. What would be the right benchmarks, datasets or tasks for productively advancing this topic?
We encourage submission of four-page abstracts describing work in progress or completed work on topics (1) and (2) above, including work from related areas, such as:
- Transfer learning
- Multi-task learning
- Meta learning
- Lifelong learning
- Few-shot learning
Finally, we will also encourage presentation of both novel approaches to CL and implemented systems, which will help concretize the discussion of what CL is and how to evaluate CL systems.
NOTE: We have available a few registrations for accepted submissions.
Razvan Pascanu (DeepMind)
Razvan Pascanu is a Research Scientist at DeepMind in London. He obtained his Ph.D. at Universite de Montreal, working with Yoshua Bengio on optimization, recurrent models and deep learning in general. His interests range from topics like optimization, neural networks to deep reinforcement learning, continual learning and structured neural networks models.
Yee Whye Teh (Oxford & DeepMind)
I am a Professor of Statistical Machine Learning at the Department of Statistics, University of Oxford and a Research Scientist at DeepMind. I obtained my Ph.D. at the University of Toronto (working with Geoffrey Hinton), and did postdoctoral work at the University of California at Berkeley (with Michael Jordan) and National University of Singapore (as Lee Kuan Yew Postdoctoral Fellow). I was a Lecturer then a Reader at the Gatsby Computational Neuroscience Unit, UCL, and a tutorial fellow at University College Oxford, prior to my current appointment. I was programme co-chair of ICML 2017 and AISTATS 2010, and gave the Breiman Lecture at NeurIPS 2017. I am interested in the statistical and computational foundations of intelligence, and works on scalable machine learning, probabilistic models, Bayesian nonparametrics and deep learning.
Marc Pickett (Google)
Marc Pickett obtained his Ph.D. at the University of Maryland under Tim Oates, working on unsupervised theory formation from uninterpreted sensors, and his M.S. under Andrew Barto at the University of Massachusetts, working on autonomous option discovery in Reinforcement Learning. He is currently at Google AI where he works on issues pertinent to long-lived systems, including transfer learning, episodic memory, and analogical schema induction.
Mark Ring (CogitAI)
Dr. Mark Ring’s research career has focused on trying to answer the following question: How should an artificial agent begin the unending process of learning and development, so that it is constantly and autonomously improving its ability to comprehend and interact with the world? His 1994 doctoral dissertation at the University of Texas at Austin, Continual Learning in Reinforcement Environments, introduced the idea of Continual Learning and began its exploration, which he continued in dozens of subsequent publications. Mark is CEO and cofounder of Cogitai, Inc., a company dedicated to bringing continual learning to the real world.