Topic: recurrent neural networks (RNNs)

The process of learning requires the brain to generate new activity patterns, yet it has been known for some time that the learning speed of certain behaviors can vary significantly. Brain computer interfaces (BCIs) offer a promising setup to study learning since they allow one to concretely link changes in activity to task performance. Using a BCI-driven center-out task, Sadtler et. al. (2014) showed that the speed of learning new activity patterns can differ significantly depending on their relation to Rhesus macaque’s intrinsic neural activity. A mapping between how individual synapses adju

Recurrent Neural Networks (RNNs) are a popular computational tool in neuroscience. Their architecture is inspired by biology, namely the abundance of lateral connections in cortex. Because they are universal approximators, they can be trained to match behavior in common neuroscience tasks, and their hidden state can be compared to neural recordings. Despite the biological inspiration, many features of real neurons are abstracted away (e.g. morphology, neuromodulators, excitability dynamics, short term plasticity). While many of these properties were included as add-ons to standard RNN models,

Behavioural tasks requiring context-dependent selection and accumulation of noisy sensory evidence are a powerful tool to study flexible decision-making (Mante et al., 2013), and recent work revealed a surprisingly degree of variability in the underlying neural mechanisms across individual animals (Pagan et al., 2022). Here we study how this individual variability develops during learning in rats and recurrent neural networks (RNNs). First, we show that the number of hidden units in RNNs plays a crucial role in the resulting amount of variability across networks initialized with different rand