Microsoft has found a new use smile – now it is a tool for training neural networks. When the engineer smiles, the system understands that all is well.
In fact, it is quite a serious study for many complex neural networks an important problem is the evaluation system of the results obtained. Usually one algorithm is responsible for the solution of the problem (e.g., creating images), and the second for its evaluation. According to experts, it is a component of the neural network is one of the most problematic.
With his assessment of the man is easier, but the question is, how real-time reading of the degree of approval of the interim solutions. The answer and the idea was to teach the algorithm to recognize some basic human emotion, which oversees the work of the neural network and this emotion naturally expresses his opinion.
Tested this method on one of the classic tasks for neural networks: a way out of the labyrinth. The output had to find the machine, which was ruled by the neural network. And during the process of watching people. If he smiled, the neural network is understood that moving in the right direction.
the Researchers say that this method of reinforcement learning was the most effective compared to any other artificial control systems. They note that it is very similar to how learning happens in real life. Children, when learning to do something, to a large extent rely on the emotions of those who teach them.
If this technology will develop, then we may train household robots in this way: by exposing them to positive or negative emotions.