Postdoc Pavel Tolmachev has a new paper out in Nature Machine Intelligence showing that the choice of single-unit activation function in a recurrent neural network (RNN) can determine the types of solutions the network can learn when trained on tasks. RNNs are commonly used to model neural dynamics, and this result highlights the importance of choosing biologically realistic activation functions so that RNN models accurately reflect how the brain solves tasks.

Congratulations, Pavel!