A DeepMind research group conducted a comprehensive generalization study on neural network architectures in the paper ‘Neural Networks and the Chomsky Hierarchy’, which investigates whether insights from the theory of computation and the Chomsky hierarchy can predict the actual limitations of neural network generalization.

While we understand that developing powerful machine learning models requires an accurate generalization to out-of-distribution inputs. However, how and why neural networks can generalize on algorithmic sequence prediction tasks is unclear.

The research group performed a thorough generalization study on more than 2000 individual models spread across 16 tasks of cutting-edge neural network architectures and memory-augmented neural networks on a battery of sequence-prediction tasks encompassing all tiers of the Chomsky hierarchy that can be evaluated practically with finite-time computation.

Using reinforcement learning (RL) to train robots directly in real-world environments has been considered impractical due to the huge amount of trial and error operations typically required before the agent finally gets it right. The use of deep RL in simulated environments has thus become the go-to alternative, but this approach is far from ideal, as it requires designing simulated tasks and collecting expert demonstrations. Moreover, simulations can fail to capture the complexities of real-world environments, are prone to inaccuracies, and the resulting robot behaviours will not adapt to real-world environmental changes.

The Dreamer algorithm proposed by Hafner et al. at ICLR 2020 introduced an RL agent capable of solving long-horizon tasks purely via latent imagination. Although Dreamer has demonstrated its potential for learning from small amounts of interaction in the compact state space of a learned world model, learning accurate real-world models remains challenging, and it was unknown whether Dreamer could enable faster learning on physical robots.

In the new paper DayDreamer: World Models for Physical Robot Learning, Hafner and a research team from the University of California, Berkeley leverage recent advances in the Dreamer world model to enable online RL for robot training without simulators or demonstrations. The novel approach achieves promising results and establishes a strong baseline for efficient real-world robot training.