For over 70 years we have told computers what to do step by step in a program. However, now, with machine intelligence, computers are starting to learn from data. This represents the biggest shift we have seen in computers since the birth of the microprocessor or perhaps since the arrival of electronic computers themselves.
In just the last few years we have seen massive advances in machine intelligence systems. Using advanced machine learning we can now accurately recognize images and we are building models that can understand the structure of language. These systems are already having a massive impact. However, we are just at the very beginning. New types of machine intelligence models are possible which deliver much higher accuracy and efficiency for perception and which will also help us build machine intelligence systems that can outperform humans at specific tasks. Transformer based Attention networks such as BERT, for example, will need higher efficiency approaches such as leveraging sparse model structures in the future so that they can more efficiently scale to billions or trillions of parameters, which will be required for richer language understanding. We will need to develop techniques that will allow us to build machines that are able to learn from experience.