Engineers have introduced an evolution system into the technology of artificial intelligence. This was reported by Science Mag with reference to Risto Miikkulainen from the University of Texas at Austin.
Learning artificial intelligence systems requires a fairly large amount of time. In neural networks, artificial neurons rather weakly imitate the structure of the real brain, mainly learning from training data. Smaller subcircuits of neurons at this time perform specific tasks – for example, to find the right traffic sign. Often, engineers spend months to connect these systems and set up these systems to work properly.
In recent years, engineers have accelerated these processes by automating certain stages of training. However, these systems are still based on a combination of ready-made algorithms developed by people.
Engineer Risto Miikkulainen and Google employee Kuok Le presented the AutoML-Zero program, which could develop artificial intelligence programs almost without human intervention, using only basic mathematical concepts.
A program can create algorithms by combining them from 100 different mathematical operations. Then the program on a simple task checks whether such an artificial intelligence system can work and compares their performance.
The parts of the different algorithms that work best will later be combined into one artificial intelligence, which will be the best variation of the algorithms.