Deep feedforward neural network models of vision dominate in both computational neuroscience and engineering. However, the primate visual system contains abundant recurrent connections. In this thesis, we investigate the addition of recurrent connections to the popular framework of convolutional neural networks for object recognition.

We train recurrent convolutional neural networks (RCNNs) on a range of small- and large-scale tasks to test for practical benefits of recurrence in object recognition. We find that recurrence improves object recognition performance relative to parameter-matched feedforward control models. Furthermore, recurrence naturally endows networks with the ability to trade-off speed and accuracy, enabling a single recurrent network to match the performance of multiple feedforward networks at a similar computational cost.

In addition, we compare the representations and connectivity of these models to the primate visual system. Firstly, the connectivity that emerges from task learning in these networks resembles the primate visual system. We also demonstrate methods for deriving reaction times from RCNNs and show that these reaction times mirror human object recognition behaviour. Moreover, RCNNs – in comparison to feedforward models – are better able to capture the overall representational dynamics of the human ventral visual pathway during object recognition.

Overall, this work shows how we can take inspiration from biology to improve performance in computer vision tasks and supports moving towards recurrent task-performing models to better understand vision within the brain.