people are fascinating.
we think, feel, act, and live as individuals in complex societies, enabled by our extraordinary cognitive abilities.
i study the neural foundations of high-level cognition, from memory to decision-making.
specifically, my current research focuses on network models of structure and reinforcement learning.
learning sensory representations for flexible computation with recurrent circuits
zhou, menendez, & latham
‧ ‧ ‧
trained rnns can perform all sorts of behaviorally relevant tasks, such as the ready-set-go task. we show here that randomly initialized untrained reservoirs can also perform such tasks if an additional linear mapping is learned between the input and recurrent layers. this drastically reduces the number of parameters that need to be tuned.
generalized energy based models
arbel, zhou, & gretton
iclr 2021 / arxiv
‧ ‧ ‧
we combine desirable properties of ebms and gans. we introduce the kale, which is a lower bound on the kl that is smooth and can be estimated from samples. after training a gan with kale, we can perform langevin sampling on the output of the generator to produce higher-quality images than those obtainable from the generator alone.
faulty towers: a hypothetical simulation model of physical support
gerstenberg, zhou, smith, & tenenbaum
cogsci 2017 [talk] / paper visualization
‧ ‧ ‧
we model people's judgments of physical stability using a noisy physics engine that can simulate possible counterfactuals. we ran several mturk experiments and confirmed that our approach better modeled intuitive human judgments.
methods of 3d printing micro-pillar structures on surfaces
ou, cheng, zhou, dublon, & ishii
uist 2015 / paper patent
‧ ‧ ‧
i wrote a java processing suite to model the geometry of 3d-printed high-resolution micropillar structures, which can collectively bend to represent a vector field. using a prepared image as input, we output a representation of the resulting structure based on image gradient properties.
[ cities ]
[ landscapes ]
[ etc ]