DeepMind and Machine learning

Google has been on the cutting edge of A.I development for some time and recently had a major advancement. Google’s DeepMind research center, consisting of a team of machine learning researchers, computational neuroscientists, and software engineers based in London, recently crafted a program that, without feeding any information to it from a human brain, learned how to play Atari games.

Machine learning has been ground breaking since its development- the ability to teach technology how to operate and learn. Educating machines has been akin to that of an infant; we have “raised” them to adopt our language, and learn from patterned experiences. Inputting information into machines over a long period of time enables the machine to pick up on patterns to eventually be able to execute on its own. But the “on its own” part has, up until recently, only been possible with human help – programing the machine with information in the effort for the technology to create relevant algorithms. However, Google’s DeepMind labs has crafted a type of machine learning program that does not require excessive inputted data from humans to effectively learn.

How is this possible?

Combining machine learning and systems neuroscience to produce general algorithms, DeepMind’s A.I. Program, coined “Deep Q-Network” (DQN), recently taught itself  how to play arcade games on its own.  Not only did DQN teach itself, the program was also able to  established farsighted strategies. It was reported that the A.I agent was capable to function right out of the box, given only “raw screen pixels, the set of actions available, and game score.”

DQN works from a combination of deep neural network and reinforcement learning – through seeking to find the familiarity and balance between the unknown, and that which is known. Reinforcement learning, inspired by behaviorist psychology, is unlike most supervised learning in that it does not correct weak actions or define the right input or output pairs. Rather, it allows software and machines to find the best context specific behavior automatically, to exploit its performance. It is trained in a way similar to how our brain experience and realize dreams and flashbacks – trained from stored samples of information taken directly from the learning phase. DQN is the attempt to create a supercomputer, a program that replicates the human brain.

Future of Machine Learning

DeepMind was initiated out of concerns over the future dangers of AI. It prioritizes the importance of human nuances, which are believed to be too complex for technology to ever understand entirely. However, the evolution of machine learning currently serves as an excellent asset to numerous fields.

A.I. has proven beneficial for certain diagnostics, especially in the medical world, where opinions and outcomes drawn from a large pool of continuously updated, world-wide data has been a great tool for providing the most relevant up-to-date information to patients. It generates global statistics from analyzed data that can then referenced to support a patient’s diagnosis. Perhaps machine learning will eventually be replacing the scientific method.

What lies ahead for machine learning is still unknown.  Perhaps leading machines will be replacing the researcher – learning as they go, coming out with the most relevant information that further helps in an endless amount of fields (stocks, economics, academia, health, etc..). Providing not only data, but data along with solutions. However, the emotional decision making and illogical tendencies of human choices is still too abstract to be captured in a program. The more advanced machine learning  that researchers are looking to develop would require complex algorithms and memory span that is, at the moment, unimaginable.