There are rich structures in off-task neural activity. In rodents, cells in the hippocampus that fire at a given place (i.e., place cells) during spatial navigation also fire spontaneously during rest, recapitulating past or future experience, referred to as neural replay. In human neuroscience, it has long been assumed that such spontaneous neural activity is task irrelevant. In fact, whole-brain activity pattern during rest is classically termed “default mode” or “task negative”. My research focus on linking this spontaneous neural activity (e.g., replay) to computational process of cognitions, and their implications to psychiatry. To achieve this, I exploit a convergent multi-disciplinary approach involving quantitative behaviour, EEG/MEG/fMRI neuroimaging, machine learning, and theory-driven computational modelling.
A key feature of my work is that it represents a strong synthesis of theory, experiment, and technical advances. In this talk, I will discuss three lines of my research on replay and cognition. Each strand supports the other, ranges from method development to translational research in psychiatry, all through the lens of computation. In short, I will introduce a novel non-invasive replay detection method in humans, enabling me to show that human replay has strong parallels to rodent replay. Moreover, using a combination of reinforcement learning theory and computational modelling, I will show that human replay during rest supports many on- task forms of cognition, e.g., represents structural knowledge explicitly for rapid inference, solves credit assignment problem for model-based learning, and supports episodic memory retrieval. When replay goes awry, for example, in patients with schizophrenia, I will show that diminished neural replay during rest is accompanied with deficits in their internal models of the world (i.e., cognitive map), suggesting a role of aberrant replay in explaining hallucination and disorganized thoughts.
In the end, I will discuss my future plans in 1) measuring sleep replay, 2) quantifying on-task replay in model-based sequential planning and 3) combining human and non-human primates (NHPs) to study complex decision-making and learning process to inform algorithm design in brain-like computation.
Yunzhe Liu is a postdoctoral research fellow at University of Oxford and University College London (UCL), working with Prof. Tim Behrens FRS, and Prof. Steve Kennerley on integrative & computational neuroscience (humans + NHPs). He works on model-based learning and decision-making using a combination of theoretical and behavioural/neuroimaging/electrophysiology approaches, with a particular interest in replay, reinforcement learning and Bayesian inference. Yunzhe did his PhD with Prof. Tim Behrens FRS and Prof. Ray Dolan FRS at MPC-UCL centre for Computational Psychiatry, and Wellcome Centre for Human Neuroimaging. Before that, he did his master in Beijing Normal University, working with Profs. Yuejia Luo, Shaozheng Qin, Yina Ma, and many others.