Technological giants such as Facebook, Google, Microsoft and IBM believe that AI will soon be the next most popular mobile device. Therefore, to accelerate progress in this area, Facebook wants to expand the community of companies that know and apply artificial intelligence in their work.
Facebook publishes its research in the field of artificial intelligence and machine learning, speaks at conferences, makes its software open sourse projects, all for the sake of exposing the myths about AI and accelerating its development. Today, for example, in its blog, the company released six short videos designed to help developers, scientists and ordinary people understand the most important aspects of artificial intelligence.
Introduction to AI (2 minutes and 40 seconds)
An explanation of how you can make machines think:
Machine learning (4 minutes and 17 seconds)
Machine learning uses models based on neural networks created using libraries such as Tensorflow (by the way, thanks to it, Google managed to achieve 94% accuracy of object recognition in pictures), Torch or Caffe. Such networks can be trained, instead of programming them for some specific tasks, for example, for recognizing and interpreting text, video and pictures. A neural network is a computer system created from several simple interconnected elements that process information based on their dynamic responses to external input data. By the way, we told how in just four steps to master the basic aspects of designing neural networks
Filling with data of universal algorithms of the model teaches this model to operate data without programming. So, for example, to train a model to translate from Russian into English, you need to submit input data to it, until the accuracy of the translation is high enough.
Gradient descent (2 minutes and 30 seconds)
Gradient descent is a method of finding a local extremum (minimum or maximum) by moving along a gradient. The following video details this method:
Deep training (1 minute and 3 seconds)
In classical machine learning, the algorithm trains on a marked sample, while the advantage of in-depth training in self-training artificial intelligence.
Method of back propagation error (1 minute and 42 seconds)
The idea is to spread the error rate (calculated in the learning step) used to recalculate the link weights in the opposite direction across all layers of the network.
Convolutional neural network (1 minute and 50 seconds)
The neural net is a special architecture of artificial neural networks, created for effective image recognition.
Also we can advise you to read a series of posts from the blog Adam Geiji (Adam Geitgey)called “Machine Learning Is Fun!”