π§ Brains π¨ Beat π€ AI at πββοΈ Fast Learning π Due to πΊ Structured Exploration π
Our biological π§ brains have an advantage over π€ AI when it comes to π rapid learning, all because of the way πΎ animals explore their environment π²π³! By understanding the brain's learning π algorithms, scientists hope to develop AI agents that can learn π faster with less experience π‘.
In a recent study π, researchers found that animals' πΎ directed exploration π is more efficient than the random π exploration shown by AI agents. Animals can learn an environment π³ in less than π minutes, while AI agents need thousands of experiences π to do the same. This discovery π emphasizes the need for better learning algorithms that can replicate the behavior of animals when exploring πΊ and learning their surroundings π.
At the Sainsbury Wellcome Centre π’ and Gatsby Computational Neuroscience Unit at UCL, neuroscientists π¬ discovered that animals' exploratory actions πββοΈ, like darting quickly towards objects, are purposeful π― and help them learn their environment efficiently π. In a paper π published in Neuron, the researchers tested whether simply observing π obstacles was enough for animals to learn about them, or if these purposeful actions were necessary for building a cognitive map πΊ.
When animals were prevented π« from performing exploratory runs πββοΈ using optogenetic tools π¦, they didn't learn their environment, even if they spent a lot of time β observing and sniffing π obstacles. This finding π shows that the instinctive exploratory actions themselves help animals learn a map πΊ of their surroundings π.
By studying different reinforcement learning models π‘ used in artificial agents, the researchers found that a combo π€ of model-free and model-based algorithms could potentially explain animal behavior πΎ. Although this may not be how the animal brain works π§ , it helps scientists understand what's required in a learning algorithm π.
The next steps for the research team π¬ include investigating the connection π between the execution of exploratory actions πββοΈ and the representation of subgoals π― in the brain. They will carry out brain recordings π§ ποΈ to identify which areas are involved in representing subgoals and how exploratory actions lead to the formation of these representations πΊ.
