I investigate how cognitive maps are formed in human brain, how neural representations of the maps change as learning goes. Previous studies have shown that 2-D map-like representation could be learned in brain regions like entorhinal cortex and hippocampus, while prefrontal cortex tend to compress the representation to 1-D. During my PhD, I use behavioral experiment, fMRI, EEG, computational models, RNNs to investigate the change of neural geometry during this learning process, what factor drives the HPC and PFC to form different neural representations, and how these learned neural representations support flexible inference and generalization.

LLMs like Centaur have shown incredible performance in modeling human behavior. However, giant need of data makes it hard to fine-tune LLMs for predicting and explaining human behaviors, and this might be the case that LLMs take short-cut to only capture human choice preferences instead of comprehensing the task structure. In order to solve this problem, we training a Transformer based on our CAMP large-scale behavioral and neural dataset collected when I was in Professor Gui Xue's lab. CAMP includes over 1,100 participants performing over 70 cognitive tasks with corresponding fMRI scannings, we believe a comprehensive dataset like this provides us possibility of building a transformer that capture human intelligence structure.

According to Wu et al., (2024), the motivation of human altruism behavior is driven by a combination of seven different socioeconomic motives, referred to as a motive cocktail. In collaboration with Xiaoyan, we use the EEG and MEG data recorded during participants performing the same task as previous behavioral study, to decode the temporal and neural neural dynamics that support each different types of motives.

Artificial neural networks are super predictive of human behaviors compared to classical cognitive models, however, the huge amount of parameters makes it hard to interpret the algorithms of mind. Thanks to HybridRNN, the framework raised by DeepMind, we are able to combine the precision of ANNs and interpretability of cognitive models to understand human learning process. Here, I have been working with Maria to apply HybridRNN to understand the computational process that enable flexible adaptations in reversal learning context, and using dynamical system analysis to understand the internal representations of HybridRNN.

With EBKernel, I am working on next generation Visual Language Navigation system based on scientific findings from human cognitive maps. Based on Tolman-Eichenbaum Machine and Vector-Hash models, human entorhinal-hippocampal circuit is an effective framework with predefined spatial grid backbone and effective item-space binding learning process, enabling human generalize their spatial and nonspatial knowledge in different envoronment. Therefore, we aim at building the best brain-inspired VLN system with high flexibility and generalizability!