When crossing a road, you want to make sure that there is no car approaching. How could an animal’s visual system detect looming events? A specific neural circuit in the fly’s visual pathway provides a solution. This neural circuit is comprised of roughly 200 neurons, called LPLC2 neurons. The dendrites of an LPLC2 neuron have four branches, with each branch extend from the cell center outwardly toward one of the four cardinal directions in space, left, right, up and down. This elegant structure emerges automatically when a neural network model is trained to perform simple looming detection tasks.

The world is always in motion and we are used to it. We see people moving around, cars passing by, trees swaying in the wind, etc.. These seemingly effortless visual experiences arise from the activities of the brain, or an intricate network of neurons. I try to understand this phenomenon using neural network models and task-oriented optimizations.

How flies navigate through odor plumes with different statistics to find food? Neural network modeling reveals the underlying algorithms.

When you are learning tennis, the coach always breaks the full swing into small steps. Similarly, it seems that songbirds also prefer learning motor skills in small steps. Our non-Gaussian Bayesian inference model provides a description of this phenomenon.