One of the most difficult challenges facing artificial intelligence today is to make laboratory-designed robots work smoothly in the real world. We see it when we try to have a fluid conversation with voice assistants like Siri: it is still bad enough to speak coherently when you have to go beyond fulfilling orders. In a research environment, the robot can be equipped with sensors and provide an ideal learning environment. But in the real world, using the same sensors would be too expensive and hostile for consumers. A group of researchers from the University of Amsterdam has set out to solve this problem.
To achieve this, they have resorted to a type of machine learning known as transfer learning. It is the process of taking what an algorithm has learned in one context and applying it in another, as it explains MIT Technology Review. It could be used to adapt an algorithm that controls a robot in the laboratory so that it can control that robot in the environment in which it has to operate and for which it was designed. That means that the robot could train first with the advantage of having better conditions and then exploit what it learned when in the real world, even when it only has cheap sensors and a hostile environment.
To test this idea, the researchers created a robot in a controlled context that moved through its surroundings with the help of eight proximity sensors and then with a single camera. They discovered that when the algorithm that controlled the robot used transfer learning to make decisions using only the information it obtained through the camera, it learned to navigate the room much faster than when it did not use transfer learning. It was also much faster than when he used transfer learning during training instead of making decisions.
. (tagsToTranslate) ia (t) can (t) make (t) robot (t) be (t) cheap (t) limit (t) skill (t) learning (t) transfer (t) ability (t) to use ( t) knowledge (t) obtain (t) context (t) teach (t) economic (t) perform (t) expensive