Different deep learning methods can train AI to complete tasks.

Think back to when you were a child. Do you remember the first time you saw a cat, or a picture of a cat? How many cats did you need to see before you fully understood exactly what a cat was?

You most likely only needed to see one or two before you could easily differentiate a cat from another animal. This sort of learning, which involves viewing a very small number of examples, comes easily and naturally to humans – but for an AI, it’s very difficult. In order for an AI to understand what makes a cat a cat, it must be trained. Nowadays, that often happens through deep learning, a form of machine learning that has driven most of the major strides in AI over the last decade.

The key message here is: Different deep learning methods can train AI to complete tasks.

Whether an AI is being trained to recognize cats, dogs, or coffee cups, it all starts with a neural network. This is software with multiple layers of “neurons” that mimic the ones found in the human brain.

There are a few different methods scientists commonly use to train neural networks. One is supervised learning, a type of deep learning in which an AI is fed a set of training examples, each labeled with a description. After the AI has been trained, we could then show it a picture of a cat. Next, the collection of pixels in the picture flows through the neural network, after which the machine will confirm – we hope – that what it “sees” is, indeed, a cat.

Even if it guesses correctly, this AI still has absolutely no idea of what the word “cat” signifies – it doesn’t know what a cat does or whether it’s alive. For an AI to develop this understanding, it needs to be taught via grounded language learning. This is a deep learning approach where sentences or words are associated with images, videos, or objects in the real world.

All of these techniques enable deep learning to have all sorts of potential applications. For one, grounded language learning could help develop AI’s language skills, making it useful in personal assistants like Siri. And deep learning has already been used to train AI to play games. In one of the most famous instances, the AI AlphaGo was trained by observing many games of Go – and ultimately was able to beat the best human champion at his own game.