Computer scientists are finding ways to code curiosity into intelligent machines, making them more efficient. Although extrinsic motivation is well-known for machine learning (i.e. reinforcement learning), limits to this approach suggest the need for intrinsic motivation too.
One way to accomplish this is by programming the machine to explore unfamiliar states in its environment. Pulkit Agrawal and Deepak Pathak, computer scientists at UC-Berkeley, have used intrinsic motivation techniques to program systems to be rewarded for being surprised.
Therefore, their curiosity is reinforced, and they are motivated to expose themselves to situations that are new and that surprise them. This allows machines to learn and explore more efficiently and flexibly — more like humans and animals.
Read the full article from Quanta Magazine here.