No abstract

Neuroscientists apply a range of common analysis tools to recorded neural activity in order to glean insights into how neural circuits implement computations. Despite the fact that these tools shape the progress of the field as a whole, we have little empirical evidence that they are effective at quickly identifying the phenomena of interest. Here I argue that these tools should be explicitly tested and that artificial neural networks (ANNs) are an appropriate testing grounds for them. The recent resurgence of the use of ANNs as models of everything from perception to memory to motor control stems from a rough similarity between artificial and biological neural networks and the ability to train these networks to perform complex high-dimensional tasks. These properties, combined with the ability to perfectly observe and manipulate these systems, makes them well-suited for vetting the tools of systems and cognitive neuroscience. I provide here both a roadmap for performing this testing and a list of tools that are suitable to be tested on ANNs. Using ANNs to reflect on the extent to which these tools provide a productive understanding of neural systems-and on exactly what understanding should mean here-has the potential to expedite progress in the study of the brain.

Artificial neural systems trained using reinforcement, supervised, and unsupervised learning all acquire internal representations of high dimensional input. To what extent these representations depend on the different learning objectives is largely unknown. Here we compare the representations learned by eight different convolutional neural networks, each with identical ResNet architectures and trained on the same family of egocentric images, but embedded within different learning systems. Specifically, the representations are trained to guide action in a compound reinforcement learning task; to predict one or a combination of three task-related targets with supervision; or using one of three different unsupervised objectives. Using representational similarity analysis, we find that the network trained with reinforcement learning differs most from the other networks. Through further analysis using metrics inspired by the neuroscience literature, we find that the model trained with reinforcement learning has a sparse and high-dimensional representation wherein individual images are represented with very different patterns of neural activity. Further analysis suggests these representations may arise in order to guide long-term behavior and goal-seeking in the RL agent. Our results provide insights into how the properties of neural representations are influenced by objective functions and can inform transfer learning approaches.

No abstract

No abstract