Maria Eckstein, Ph.D., Google Deep Mind

Topic: Understanding Human Learning and Decision Making Using Cognitive Models and Artificial Neural Networks

Computational modeling has been an indispensable tool for cognitive science, offering insights into otherwise elusive cognitive mechanisms. From Drift Diffusion Models to Reinforcement Learning (RL), Bayesian Inference, and beyond, computational models have not only illuminated intricate cognitive processes, but also paved the way for a deeper understanding of the underlying neural substrate. In this presentation, I will first discuss a study that uses the RL framework to provide a precise, mechanistic explanation of how the ability to learn from reinforcement develops in humans. My focus will be on the adolescent years, marked by tremendous changes both to the external environment and to neural substrates relevant to learning. Subsequently, I will address my research on some inherent limitations of classic cognitive models and present a recent project at the intersection of cognitive psychology and artificial intelligence that aims to overcome these challenges. I will introduce a model that augments classic RL models with artificial neural networks, and show how this method allows us to paint a clear, yet comprehensive picture of human reward-based learning. Deriving its computations directly from observed behavior, my model obliterates the need to hand-specify equations, and offers superior fit to human behavior. At the same time, the model offers improved interpretability: It reveals the necessity for adaptable cognitive processes, and shows that a strong and flexible memory system needs to be at the core of our understanding of learning, enabling representation learning capabilities that go far beyond classic cognitive models like RL. Overall, this approach highlights the need - and feasibility - of embracing more complexity in cognitive theories. In the penultimate section of my talk, I will pivot to discuss my research in one suc area: abstraction. I will briefly touch on several topics, from the time course of rule-based hierarchical inference, to a hierarchical RL model explaining human learning across contexts. In the final part of my talk, I will outline my plans for future work, including the use of procedurally-generated task paradigms, and the application of neural-network-based cognitive models in the analysis of neural data.

About Maria Eckstein: I am a senior research scientist at Google DeepMind with a PhD in psychology from UC Berkeley and a B.A. in philosophy and a B.Sc. in psychology from Ludwig Maximilian University in Munich, Germany.

My work lies at the intersection of AI and cognitive science / neuroscience. I combine tools from artificial intelligence, such as reinforcement learning and neural networks, with those from cognitive psychology and neuroscience, such as controlled lab experiments and moment-to-moment neural recordings. I am particularly interested in questions that lie at the intersection of AI and cognitive science / neuroscience, including learning, decision making, and structured thought.