Understanding human metacontrol and its pathologies using deep neural networks

What is it about?

Computational systems have to learn when and how they should exert control over their actions. How do agents learn to solve this “metacontrol” problem? Here, we created a task that externalizes this control (through special actions that sample more information about the world or increase efficacy). Using deep neural network models, we show that only agents equipped with brain modules that explicitly compute the likely success of their actions perform optimally, behaving very similarly to human participants. This helps us understand why mammals have evolved brain signals, in the medial prefrontal cortex, that compute the likely success or failure of their actions. Moreover, perturbing this computation leads to behaviors reminiscent of psychological disorders such as depression, anxiety, and compulsion.

Featured Image

Photo by Bhautik Patel on Unsplash

Photo by Bhautik Patel on Unsplash