Imagine being in a new town, surrounded by unfamiliar sights and sounds. Despite the initial disorientation, you gradually create a mental map of your surroundings, understanding where buildings, streets, and signs are in relation to each other. This ability to construct spatial maps in the brain is a fundamental aspect of human cognition.
According to Matt Thomson, an assistant professor of computational biology, cutting-edge artificial intelligence and neural networks still fall short in this area. While AI models can perform tasks based on rote memorization, they struggle to problem-solve and generate new ideas like humans do. Thomson believes this is because they lack the ability to navigate in conceptual space, moving through a space of ideas much like navigating physical space.
Breakthrough in Neural Network Spatial Mapping
A recent study led by graduate student James Gornet at Caltech’s Department of Computational and Neural Systems has made significant progress in this area. Using a predictive coding algorithm, the team trained a neural network to build spatial maps within the game Minecraft. By recording videos of a player exploring the game environment and analyzing the data, they found that the neural network could predict upcoming environments and store representations of objects spatially within its structure.
This breakthrough represents the first time a neural network has been shown to create its own map, rather than relying on pre-existing data provided by human designers. The ability to spatially organize information could lead to advancements in AI, enabling neural networks to solve complex problems independently.
Funding and Future Implications
The study, titled “Automated construction of cognitive maps with visual predictive coding,” was funded by organizations such as The David and Lucile Packard Foundation and the Chan Zuckerberg Initiative. Thomson emphasizes the importance of the bioinspired approach to machine learning, which aims to reverse-engineer properties of the brain in artificial neural networks. This innovative work at Caltech’s CNS program showcases the potential for interdisciplinary research to drive progress in the field of AI.
