AI Learning About Real World by Playing Video Games
Artificial intelligence is learning how to interact with the real world by playing video games...just like most young people. Joking aside, the new video game based training techniques are highly valuable, especially for smaller AI companies, and should improve the reliability and safety of all autonomous systems.
"Deep learning" algorithms, that form the foundation of modern artificial intelligence, sift through large datasets to find patterns. Dataset creation and annotation is costly and takes a lot of time to perform. First, in the example of autonomous cars, the AI trainer must collect a dataset by driving around and recording the types of scenes an autonomous vehicle will see while on the road. Next, each frame of the video must be analyzed by technicians to identify and annotate the elements of the image that are road, pedestrian, building, etc so that the AI can learn to distinguish each element. The annotating of each frame is time consuming and constrains the total amount of footage that can be fed into the algorithm.
However, video game engines can automatically annotate each element it generates in its virtual world, saving AI trainers from having to review each frame. This annotating capability, along with the ability of a video game engine to generate a near infinite range of atypical conditions such as unusual weather, unusual lighting, unusual obstacles, etc., is highly valuable.
Videos describing how video games are used to train autonomous automobiles and autonomous aircraft are provided below.
Why this news is important
Autonomous systems are here to stay but the reliability and safety of these systems must be extremely high in order to avoid regulations that will stifle innovation. Video games can be used as a cost effective method to train an artificial intelligence about the vast range of scenarios it may face in the real world.
Image Credit and References
Product website, Neurala, Collision Avoidance Training for Drones, September 21, 2016.
