How to make a neural network smarter

Neuroscience experiments usually ask participants to make one simple choice: left or right, red or blue. But the real world is far more complex: imagine walking down a fruit aisle, deciding both which berries are ripe and which berries would work in the recipe you have planned.

In a new Nature Communications paper, Zuckerman Institute's Stefano Fusi and Jeff Johnston show that neural networks that are required to multitask develop more abstract representations of the world.

These abstract representations may underlie “a remarkable ability to generalize knowledge across distinct contexts and objects.”