AI Advances in Problem-Solving by Mapping Minecraft Spaces

A neural network designed to build spatial maps using a predictive coding algorithm trained with Minecraft demonstrated by recent research from Caltech’s Thomson lab.

Recent research led by the Thomson lab at Caltech has revealed impressive new advances in artificial intelligence (AI), focusing on how neural networks can create spatial maps. This ability, crucial for advanced cognitive functions in humans, has now been demonstrated in AI using predictive coding algorithms and somewhat surprisingly. Minecraft.

Humans can navigate unfamiliar environments and create mental maps, a skill that supports complex cognitive functions such as language processing. While AI has made significant strides, it has not yet matched this human capability. According to Matt Thomson, assistant professor of computational biology and Heritage Medical Research Institute Investigator, “There’s this sense that even state-of-the-art AI models are still not truly intelligent. hey don’t problem-solve like we do; they can’t prove unproven math results or generate new ideas. We think it’s because they can’t navigate in conceptual space; solving complex problems is like moving through a space of concepts, like navigating.”

Thomson’s team, including graduate student James Gornet, addressed this limitation by developing neural networks capable of constructing spatial maps. They conducted experiments within the game Minecraft, designing complex environments with elements like trees, rivers, and caves. By recording videos of a player moving through these environments and training a neural network with predictive coding algorithms, they demonstrated the network’s ability to learn the spatial organisation of objects within the game world. The network could predict upcoming environments during navigation, suggesting it had formed a spatial map.

Moreover, when the researchers examined the neural network’s internal structure, they found that it stored representations of objects spatially in relation to each other, effectively creating a map of the Minecraft environment. This achievement marks the first instance of a neural network independently generating a spatial map, a critical step towards more sophisticated AI problem-solving abilities.

Traditional neural networks can navigate maps provided by human designers, such as self-driving cars using GPS. However, the capability to create their own maps opens new possibilities for AI to tackle complex problems akin to human problem-solving. As Gornet, a student in Caltech’s Department of Computational and Neural Systems (CNS), explains, “The CNS program really gave James a place to do unique work that wouldn’t be possible anywhere else.” The department’s interdisciplinary approach, encompassing neuroscience, machine learning, mathematics, statistics, and biology, fosters innovative research that bridges biological insights with artificial neural networks.

Thomson adds, “We’re taking a bioinspired approach to machine learning that allows us to reverse-engineer properties of the brain in artificial neural networks, and we’re hoping to learn about the brain in turn. We have a very receptive community for this kind of work here at Caltech.

It would be interesting to see how much more we could understand if the model was trained on a Minecraft parkour map, or perhaps the scale Minecraft model of CERN.

You can read the full paper “Automated construction of cognitive maps with visual predictive coding” in the journal Nature Machine Intelligence.