How can deep learning models flexibly reuse their knowledge? How can they adapt to new tasks zero-shot, as humans can? In our new preprint (arxiv.org/pdf/2005.04318), we propose a new approach based on learning to transform task representations: meta-mapping. Preview in thread:
Our approach can make drastic adaptations zero-shot, like switching from winning at (simplified) poker to trying to lose. It can allow a visual classification system to recognize new concepts, and can adapt a model-free reinforcement learning to new tasks, without data from them.
It accomplishes this without prior domain knowledge, based only on the relationships between tasks. Specifically, it learns basic task representations, e.g. for poker, via meta-learning. It also learns meta-mappings, higher order tasks which transform these basic task reps.
For example, it might learn a "lose" meta-mapping, from the relationship between winning and losing at games like blackjack. This meta-mapping could then be applied to the model's representation of poker, in order to lose at poker zero-shot.
We show this method can allow 80-90% performance, zero-shot, in domains ranging from polynomial regression to visual classification to reinforcement learning, outperforming baselines (sometimes substantially). It even exhibits some intriguing signatures of being more systematic.
This zero-shot adaptation then allows the system to master the new tasks much more efficiently. It makes an order of magnitude fewer mistakes (cumulative loss) on the way to mastering the tasks than the next-best approach we considered.
We implement this all in a parsimonious, homoiconic architecture that reuses the same networks for basic tasks and meta-mappings. This improves generalization! We also show our approach works with task representations constructed from either examples of the task or language.
I think that meta-mapping may offer a useful concept for building more flexible artificial intelligence systems, and better cognitive models.
Thanks for making it through this thread! There's a lot more detail, experiments, and related work/implications in the paper, please check it out! I hope it will be interesting and understandable to researchers in both AI/ML and cognitive science. arxiv.org/pdf/2005.04318
