Artificial Recurrent Neural Networks (ANN) such as LSTM and GRU have recently gained an intense interest in machine learning. These models are highly versatile and achieve breakthrough performance in many applications of sequential data. However, the learning algorithm of ANN, such as backpropagation, is challenging to implement in biological circuits with living cells.

In this project, we focus on an archetypal circuit of biological learning --- the mushroom body in insects. Despite its simple design, the circuit is responsible for the complex navigation behaviors in honeybees and fruit flies. Over the past decades, the field has accumulated comprehensive experimental data on the structure (connections between neurons) and function (plasticity or how the connection strength between neurons changes in response to the environment) of the mushroom body, setting a perfect stage to advance the circuit at the theoretical level.

In this project, we seek to (i) understand the key components of its learning algorithms, such as plasticity rules and types of connections between circuit layers, and (ii) characterize mathematically the capacity (i.e., how many stimulus patterns an animal can remember) of the mushroom body circuit given the components in (i).