Structured Exploration Allows Animal Brains to Learn Faster than AI
Researchers at University College London have uncovered how animals’ instinctive exploratory runs are critical in facilitating their learning and building a cognitive map of their environment. The study, published in Neuron, indicated that the actions animals take, such as lunging for objects, are key to efficient learning that allows mice to learn a map of the world quickly. The UCL team found that animals’ purpose-driven actions played an essential role in enabling their cognitive mapping of the world. Unlike artificial agents, animal exploration is targeted and focused on distinctive objects, allowing for better, more efficient learning than artificial agents.
The study’s lead author, Philip Shamash of the Sainsbury Wellcome Center at UCL, carried out experiments to test the effect of preventing animals from exploring. Through his experiments, he found that mice could not adequately learn if they could not undertake exploratory runs towards obstacles. This showed that the animals’ own instinctive actions helped them learn more efficiently.
The study also identified two classes of reinforcement learning models – modelless and model-based – that the mice use to learn new things. Under some conditions, the mice acted without a model, while in others, they had a model of the world. The researchers implemented an algorithm that arbitrates between modelless and model-based learning to explain animal behavior.
Professor Tiago Branco, group leader at the Sainsbury Wellcome Center at UCL and corresponding author on the paper, said that there was a significant difference between artificial intelligence agents and animal exploration. While agents need a vast amount of experience to learn, animals can learn an environment within ten minutes. Structured exploration has the power to help animals learn more efficiently, which is why they require less experience than AI agents to learn.
The researchers’ next step will be to record brain activity to identify which areas are involved in the representation of subgoals and how exploratory actions result in the formation of those representations.
As per the source, subscribe to Technology Networks’ daily newsletter for more breaking science news delivered straight to your inbox every day.
Reference: Shamash P, Lee S, Saxe AM, Branco T. Mice identify secondary target locations through an action-driven mapping process. Neuron. 2023: S0896627323002301. do: 10.1016/j.neurona.2023.03.034
This article has been republished as per materials from the source. Note: the material may have been edited for length and content. For more information, contact the cited source.
