Jane X Wang, Zeb Kurth-Nelson, Dhruva Tirumala
Hubert Soyer, Joel Z Leibo, Remi Munos, Charles Blundell, Dharshan Kumaran, Matt Botvinick
2016 · arXiv
In recent years deep reinforcement learning (RL) systems have attained superhuman performance in a number of challenging task domains. However, a major limitation of such applications is their demand for massive amounts of training data. A critical present objective is thus to develop deep RL methods that can adapt rapidly to new tasks. In the present work we introduce a novel approach to this challenge, which we refer to as deep meta-reinforcement learning. Previous work has shown that recurrent networks can support meta-learning in a fully supervised context. We extend this approach to the RL setting. What emerges is a system that is trained using one RL algorithm, but whose recurrent dynamics implement a second, quite separate RL procedure. This second, learned RL algorithm can differ from the original one in arbitrary ways. Importantly, because it is learned, it is configured to exploit structure in the training domain. We unpack these points in a series of seven proof-of-concept experiments, each of which examines a key aspect of deep meta-RL. We consider prospects for extending and scaling up the approach, and also point out some potentially important implications for neuroscience.
Metadata: arXiv. Abstracts retain their original authors’ rights.
Original HN comments, with labels from simple text rules. A label describes the whole comment and may miss context or irony. Read the discussion to make up your own mind. About the labels
5 selected excerpts · 5 total mentions
Hyperparameter optimization (including architectures) is not really meta-learning. Meta-learning, also known as "learning to learn", is more like MAML[0], RL2[1], L2RL[2], etc. 0. https://arxiv.org/abs/1703.03400 1. https://arxiv.org/abs/1611.02779 2. https://arxiv.org/abs/1611.05763
Many concerns in this paper, especially about deep learning being data hungry, limited capacity for transfer, and integrating prior knowledge are addressed by recent papers in meta-learning and learning-to-learn [1,2,3 and many many others] both in the supervised and reinforcement learning contexts. In the case of meta-reinforcement learning there has been recent work [5] which seems to indicate that this mechanism is very similar to how learning works in the brain. In fact, my group recently published a paper [4] about learning-to-learn in spiking networks (making it biologically more realistic) and showing that the network learns priors on families of tasks for…
supervised learning, and learns useful exploration strategies automatically when doing meta-reinforcement learning. While I don't claim this is the right path to AGI, it's a very promising and new direction in deep learning research which this paper seems to ignore. [1] http://citeseerx.ist.psu.edu/viewdoc/summary?doi=10.1.1.5.32... [2] https://arxiv.org/abs/1611.05763 [3] https://arxiv.org/abs/1703.03400 [4] https://arxiv.org/abs/1803.09574 [5] https://www.nature.com/articles/s41593-018-0147-8 (Preprint: https://www.biorxiv.org/content/early/2018/04/06/295964)
Handling delays (and the uncertainty they entail) is a huge challenge, and I think it'll be a rich area of research. The simplest part of the problem is that delays in action or perception also slow the propagation of reward signals, and credit assignment is still a really hard problem. Thinking further afield, future models could learn to adapt their expectations to fit the behavior of a particular opponent. This kind of metalearning is pretty much a wide open problem, though a pair of (roughly equivalent) papers in this direction recently came out from DeepMind: https://arxiv.org/abs/1611.05763 and OpenAI: https://arxiv.org/abs/1611.02779 It's going to be really exciting to see how these techniques scale.
Meta- reinforcement learning could prove to be such breakthrough, see [1],[2]. Also next generation ASIC accelerators (Google's TPU, Nervana) can give 10x increase in NN performance over a GPU manufactured on the same process, with another 10x possible with some form of binarized weights, e.g. BNN, XNOR-net. There are also interesting techniques to update the model's parameters in a sparse manner. So, there certainly is a lot of room left for performance improvements! 1. https://arxiv.org/abs/1611.05763 2. https://arxiv.org/abs/1611.02779
Looks like the "UNREAL" (https://arxiv.org/abs/1611.05397), "Learning to reinforcement learn" (https://arxiv.org/abs/1611.05763) and "RL^2" (https://arxiv.org/abs/1611.02779) are state of art in pure RL for now. Finally there is a trend of using recurrent neural network as a top component of the Q-network. Perhaps we will see even more sophisticated RNNs like DNC and Recurrent Entity Networks applied here. Also we'll see meta-reinforcement learning applied to a curriculum of environments.
More from the same topics.
Tom B. Brown, Benjamin Mann, Nick Ryder +28
Recent work has demonstrated substantial gains on many NLP tasks and benchmarks by pre-training on a large corpus of text followed by fine-tuning on a specific task.
Chiyuan Zhang, Samy Bengio, Moritz Hardt +2
Despite their massive size, successful deep artificial neural networks can exhibit a remarkably small difference between training and test performance.
Haoran Xu, Fredrik Kjolstad
Fast compilation is important when compilation occurs at runtime, such as query compilers in modern database systems and WebAssembly virtual machines in modern browsers.
Scott Aaronson
In honor of Alan Turing's hundredth birthday, I unwisely set out some thoughts about one of Turing's obsessions throughout his life, the question of physics and free will.