AI Weekly: Continual learning offers a path toward more humanlike AI

State-of-the-art AI systems are remarkably capable, but they suffer from a key limitation: statisticity. Algorithms are trained once on a dataset and rarely again, making them incapable of learning new information without retraining. This is as opposed to the human brain, which learns constantly, using knowledge gained over time and building on it as it encounters new information. While there’s been progress toward bridging the gap, solving the problem of “continual learning” remains a grand challenge in AI.

This challenge motivated a team of AI and neuroscience researchers to found ContinualAI, a nonprofit organization and open community of continual and lifelong learning enthusiasts. ContinualAI recently announced Avalanche, a library of tools compiled over the course of a year from over 40 contributors to make continual learning research easier and more reproducible. The group also hosts conference-style presentations, sponsors workshops and AI competitions, and maintains a repository of tutorial, code, and guides.

As Vincenzo Lomonaco, cofounding president and assistant professor at the University of Pisa, explains, ContinualAI is one of the largest organizations on a topic its members consider fundamental for the future of AI. “Even before the COVID-19 pandemic began, ContinualAI was funded with the idea of pushing the boundaries of science through distributed, open collaboration,” he told VentureBeat via email. “We provide a comprehensive platform to produce, discuss and share original research in AI. And we do this completely for free, for anyone.”

Even highly sophisticated deep learning algorithms can experience catastrophic learning or catastrophic interference, a phenomenon where deep networks fail to recall what they’ve learned from a training dataset. The result is that the networks have to be constantly reminded of the knowledge they’ve gained or risk becoming “stuck” with their most recent “memories.”

OpenAI research scientist Jeff Clune, who helped to cofound Uber AI Labs in 2017, has called catastrophic forgetting the “Achilles’ heel” of machine learning and believes that solving it is the fastest path to artificial general intelligence (AGI). Last February, Clune coauthored a paper detailing ANML, an algorithm that managed to learn 600 sequential tasks with minimal catastrophic forgetting by “meta-learning” solutions to problems instead of manually engineering solutions. Separately, Alphabet’s DeepMind has published research suggesting that catastrophic forgetting isn’t an insurmountable challenge for neural networks. And Facebook is advancing a number of techniques and benchmarks for continual learning, including a model that it claims is effective in preventing the forgetting of task-specific skills.

But while the past several years have seen a resurgence of research into the issue, catastrophic forgetting largely remains unsolved, according to Keiland Cooper, a cofounding member of ContinualAI and a neuroscience research associate at the University of California, Irvine. “The potential of continual learning exceeds catastrophic forgetting and begins to touch on more interesting questions of implementing other cognitive learning properties in AI,” Cooper told VentureBeat. “Transfer learning is one example, where when humans or animals learn something previously, sometimes this learning can be applied to a new context or aid learning in other domains … Even more alluring is that continual learning is an attempt to push AI from narrow, savant-like systems to broader, more general ones.”

Even if continual learning doesn’t yield the sort of AGI depicted science fiction, Cooper notes that there are immediate advantages to it across a range of domains. Cutting-edge models are being trained on increasingly larger datasets in search of better performance, but this training comes at a cost — whether waiting weeks for training to finish or the impact of the electricity usage on the environment.

“Say you run a certain AI organization that built a natural language model that was trained over weeks on 45 terabytes of data for a few million dollars,” Cooper explained. “If you want to teach that model something new, well, you’d very likely have to start from scratch or risk overwriting what it had already learned, unless you added continual learning additions to the model. Moreover, at some point, the cost to store that data will be exceedingly high for an organization, or even impossible. Beyond this, there are many cases where you can only see the data once and so retraining isn’t even an option.”

While the blueprint for a continual learning AI system remains elusive, ContinualAI aims to connect researchers and stakeholders interested in the area and support and provide a platform for projects and research. It’s grown to over 1,000 members in the three years since its founding.

“For me personally, while there has been a renewed interest in continual learning in AI research, the neuroscience of how humans and animals can accomplish these feats is still largely unknown,” Cooper said. “I’d love to see more of an interaction with AI researchers, cognitive scientists, and neuroscientists to communicate and build upon each of their fields ides towards a common goal of understanding one of the most vital aspects of learning and intelligence. I think an organization like ContinualAI is best positioned to do just that, which allows for the sharing of ideas without the boundaries of the academic or industry walls, siloed fields, or distant geolocation.”

Beyond the mission of dissemination information about continual learning, Lomonaco believes that ContinualAI has the potential to become a reference points for a more inclusive and collaborative way of doing research in AI. “Elite university and private company labs still work mostly behind close doors, [but] we truly believe in inclusion and diversity rather than selective elitiarity. We favor transparency and open-source rather than protective IP licenses. We make sure anyone has access to the learning resources she needs to achieve her potential.”

For AI coverage, send news tips to Kyle Wiggers — and be sure to subscribe to the AI Weekly newsletter and bookmark our AI channel, The Machine.

Thanks for reading,

Kyle Wiggers

AI Staff Writer