This is the neural network of DeepMind, a company specialized in artificial intelligence that was acquired by Google due to its merits in this field.
Just playing video games, for now. Science goes step by step. DeepMind engineers have built an artificial neural network with the ability to learn and make decisions based on the knowledge acquired. In this case it applies to video games, a field where researchers they can comfortably experience and evaluate the reactions of the system, and then compare it with those that occur in a human brain.
The tests DeepMind’s artificial brain has undergone consisted of playing a 49 different arcade games from the 70s, which included legendary titles such as Pong and Space Invanders. The system was not specifically programmed to play video games, it was simply imbued with the ability to learn. Here it is worth making a fundamental distinction, which marks the difference between artificial intelligence and an ordinary software algorithm, such as Deep Blue, which beat chess champion Gary Kasparov.
In this case, the DeepMind system starts playing and learns as you practice, just like a person would. In this way he gets to acquire such a master’s degree in video games that it is capable of beating human professionals. However, each new game requires new training. The system always starts again from scratch, because it is not able to transfer the previously acquired knowledge to the new scenario, as a person does.
Their method of learning is similar to the one a person demonstrates. The artificial brain receive stimuli – in this case the score assigned by the game in each game– and analyze your plays based on this data. Mixing these two variables results in the following decision-making.
Two basic technologies
DeepMind, which was acquired last year by Google, has based its neural network on two artificial intelligence technologies. The first is the machine learning or machine learning, which uses an architecture inspired by the brain, in which the connections between layers of simulated neurons are strengthened based on experience. In this way the system can extract complex information from unstructured fragments. A quality that Google has delved into due to the great interest it has for its online advertising business.
The second technology is called ‘reinforcement learning’. This term, which is used in psychology to define the positive reinforcement that is practiced with people to encourage them to do things well, tries to mimic the neural reward system in an animal. It is therefore a matter of reinforcing the knowledge learned to avoid making mistakes.
DeepMind’s artificial brain has among its objectives the study of the human brain. That is why, among other reasons, video games have been chosen as a field of experimentation. Because it is easier to assimilate the activity of neurons in this way, since so you can compartmentalize the study, checking how the different parts of the system work with each action of the gamer-machine.
Images: Space Invanders and thelunch_box