Recent advances in artificial intelligence (AI), particularly in deep learning, have demonstrated remarkable success across diverse domains such as vision, reinforcement learning, and natural language processing. The most advanced neural networks today are capable of performing a wide array of cognitive tasks at a human level, marking a significant milestone in AI research. Moreover, these models are becoming integral to our daily lives, further underscoring their transformative potential. 

While the neural networks powering these models are biologically implausible in many ways, they still share some core computational principles with the brain, such as the use of incremental learning and distributed representations. The success of these AI systems presents an exciting opportunity to explore fundamental principles of cognition and intelligence. Can we leverage recent advances in AI to help us understand emergent cognitive processes in humans? And conversely, can insights from neuroscience refine and enhance the capabilities of AI models? 

We are investigating these questions through ongoing collaborations with labs in the computer science department (including Professor Ellie Pavlick and the LUNAR lab). In particular, we have found that transformers trained on working memory tasks learn internal circuits resembling those implicated in the functioning of the prefrontal cortex and basal ganglia (see ref and ref). Another project has investigated how emergent in-context learning abilities in transformers allow them to replicate compositional generalization behaviors and curriculum effects observed in humans, also thought to rely on frontostriatal circuitry (see ref and ref). Additionally, an ongoing project investigates how transformers develop control strategies analogous to the proactive and reactive modes of cognitive control observed in humans when trained on simple cognitive tasks.