Training data for olfaction is scattered across disparate, non-standardized datasets, limiting the ability to build representative world models. Olfactory navigation is a highly dynamic and non-stationary task that benefits from real-time continual learning.
We introduce an adaptive framework called Grow-Prune-Freeze (GPF) networks that enables an agent to continually learn by growing, pruning, and freezing early layers of its policy in response to world complexity.
Grounding GPFs in non-linear random matrix theory, we show that the work of Pennington & Wirth (2017) can be extended from single hidden layers to n-layer continual-learning models, and the eigenvalue composition of network weights is preserved as successive layers are added.
GPFs based on Expected SARSA achieve a 94% success rate on turbulent plume navigation—a partially observable, non-stationary task representative of the real-world challenge of adaptive learning in robotics.
Furthermore, our experimental results indicate that GPF networks may generalize well to other machine learning tasks, such as reinforcement learning in Atari games, image classification, and autoregressive language modeling. We have open-sourced all code and data to encourage improvements and further research in the field of olfactory robotics.
Author's Comment: GPF networks provide a new solution for olfactory navigation, offering adaptive learning to cope with complex environments and broad application potential. This approach is not only significant in robotics but may also advance other machine learning tasks, making it worthy of further exploration.