Google is working on a better AI for smartphones

Applications that detect objects, classify objects or recognize faces are nothing new. However, they most often use neural networks, the disadvantage of which is slow operation and inaccuracy.

Scientists from Google have just presented a new approach to AI, which is much faster and more precise. It consists of three things: a recursive neural network that learns and creates sample models, a trainer that creates and trains models, and an engine that measures the speed of the models created. Google tested the models created in this way on a set of data obtained from Stanford and Princeton universities and on the basis of Common Objects in Context (COCO). It turned out that they are 1.5 times faster than the best systems based on neural networks. The cost of computing power alone is 35 times lower compared to the SSD300 benchmark.

Research is conducted offline and is constantly being developed. Many companies are working on improving AI for object recognition, but it is possible that Google has overtaken them.

Source: https://venturebeat.com/