The evolution of computers has gone a long way from vacuum tubes and ICs to artificial intelligence. Deep learning, a branch of machine learning has put in a great deal of effort to make computers think and recognize images and speech like humans. Now that it has achieved its goal, the tech giants and startups are moving towards new tools to take artificial intelligence to the next level. It is to help the computers to achieve a goal after it has recognized the problem. This needs two skill sets, memory and mimicry. Two leading tech giants, Google and Facebook have already created products using these skill sets. Facebook has revealed that its Memory Nets are helping computers learn like babies do. Google has acquired a firm called Deep Mind, a predecessor in these efforts, to improve its translation and Smart Reply services. DeepMind has worked on teaching the computer play video games on its own using artificial intelligence. They had worked so hard and for so long to teach the computer to understand what is on the screen, take random actions, analyze the statistics of the game and learn what actions made the scores get higher.
On Wednesday, a new machine learning startup named Osaro was launched with a fund of $3.3 million from Peter Thiel, Scott Banister, and Yahoo’s Jerry Yang. This startup aims to bring deep reinforcement learning from research to production. For instance, once the computer can recognize that the mail is sent to ask for an appointment, it has to consider about the possible replies without human intervention. Derik Pridmore, the co-founder of Osaro has said that the computers will need a policy to know what comes next to achieve a goal. This policy is now being devised by engineers manually. Their aim is to use deep learning for this purpose and tutor a computer that its action was right when it has achieved the goal. Pridmore has aimed to build a production system this year that would teach robots to recognize situations and decide the actions on its own. Osaro is providing software similar to the operating system that could be used in industrial robotics that would enable the robots to perform an action and learn from it. This feature of the software makes the robot to adapt to the changing environment. But achieving this feature in a robot was not easy for Osaro like it was for DeepMind. Taking random actions to learn things were fine for video games but not while training a robot as it could turn deadly. So Osaro wrote algorithms that help computers learn by mimicking humans. It is not only Osaro and DeepMind, but also the University of Washington that is aiming to teach robots.
Sumit Sanyal, the CEO of a startup called “Minds.ai” which is building hardware for deep learning has said that this type of reinforcement learning along with the memory for it would make artificial intelligence stronger in the forthcoming years.