Speaker: Chris Cueva

Title: Building and evaluating recurrent neural networks on multiple datasets

Abstract: Natural intelligence entails the interactions between many systems: perception, cognition, action, memory, etc. and so ultimately many of the open questions in systems neuroscience will require models that bridge these systems. To make progress towards these multi-system models we are creating a high-throughput pipeline for training different recurrent neural network (RNN) models on a wide range of tasks and comparing them to experimental datasets. We have been inspired by community-wide efforts using mostly feedforward networks (e.g. ImageNet and Brain-Score) centered around benchmarks to both improve model architectures and evaluate model fits to data, and felt the time is ripe for similar efforts with recurrent models that encompass a larger diversity of tasks and brain regions. In this talk I’ll share some of our initial progress towards refining and testing RNN models by evaluating them against multiple datasets.