So why not ask the neurons what they want to see?

Read: The human remembering machine

That was the idea behind XDREAM, an algorithm dreamed up by a Harvard student named Will Xiao. Sets of those gray, formless images, 40 in all, were shown to watching monkeys, and the algorithm tweaked and shuffled those that provoked the strongest responses in chosen neurons to create a new generation of pics. Xiao had previously trained XDREAM using 1.4 million real-world photos so that it would generate synthetic images with the properties of natural ones. Over 250 such generations, the synthetic images became more and more effective, until they were exciting their target neurons far more intensely than any natural image. “It was exciting to finally let a cell tell us what it’s encoding instead of having to guess,” says Ponce, who is now at Washington University in St. Louis.

This paper published in Nature on 26th February 2015, describes a DeepRL system which combines Deep Neural Networks with Reinforcement Learning at scale for the first time, and is able to master a diverse range of Atari 2600 games to superhuman level with only the raw pixels and score as inputs.

For artificial agents to be considered truly intelligent they should excel at a wide variety of tasks that are considered challenging for humans. Until this point, it had only been possible to create individual algorithms capable of mastering a single specific domain. With our algorithm, we leveraged recent breakthroughs in training deep neural networks to show that a novel end-to-end reinforcement learning agent, termed a deep Q-network (DQN), was able to surpass the overall performance of a professional human reference player and all previous agents across a diverse range of 49 game scenarios.