Artificial intelligence and machine learning have come a long way and have become buzzwords in many tech products today. Impressive as they are, however, their methods of learning are still mostly, well, artificial. NVIDIA researchers are developing a new way to train AI for industrial robots that almost closely mimics the way we ourselves learn. And that’s by watching how another, more experienced human perform a task and then trying to repeat it.
Most neural networks are trained using what almost seems like brute force, except that it takes hours rather than years. Machine learning often involves feeding a neural network thousands upon thousands of pieces of pre-existing data. While effective, it isn’t always efficient nor is it applicable in all situations.
Take, for example, factory or industry robots. They could very well benefit from some machine learning but training them often requires a specialist to program the robots at a low-level. In contrast, NVIDIA’s researchers have come up with an AI that is able to learn just by watching a human go through the task first. It even learns just by seeing it once.
Of course, there’s still some behind the scenes learning involved but even then it’s not your usual machine learning. NVIDIA prefers the use of “synthetic data” versus labeled training data, which gives the neural network an almost infinite amount of data to work with for very little effort. These simulation-based learning is, of course, well suited for NVIDIA’s GPUs, like the AI-oriented Titan X.
This new method from NVIDIA researchers is also notable for other things. For one, the AI can lay out the steps it will take in plain English so that scientists and programmers can immediately see and, if necessary, correct the steps. Such a kind of robot AI would also eventually allow machines to not only learn from human co-workers but even safely operate alongside them.