The Differences Between AI, Machine Learning, and Deep Learning

Not to steal the thunder from today’s topic, but Rob Pope, our very own CTO and Co-Founder, will be speaking at the Smart Home Summit this week in Palo Alto, CA. If you’re a developer based in San Francisco, come see Rob drop some words of wisdom about AI and IoT. Learn more here!

We recently wrote an overview of artificial intelligence (AI) that discussed its history and where it’s going next. Besides just the cold, hard facts, there are myriad innovations happening to the AI field right now. 2018 will bring the launch of an “AI app store.” And AI is completely and irreversibly transforming the world of mobile advertising right now. So, for you mobile app developers out there, it pays to know the basics.

Whether the AI you first think of is Google DeepMind’s AlphaGo, the robot who defeated South Korean Master Lee Se-dol, or Spotify’s intelligent song and playlist recommendation engine, there’s no doubt that AI is already infiltrating the world around us in silent, astonishing ways. When Google’s DeepMind team released press statements about how AlphaGo was created, they used buzzwords like “AI”, “machine learning,” and “deep learning”; all three are connected and intertwined, but at the same time, each is its own field.

An easy way to think about the relationship between the three fields is to imagine three circles inside of each other, each one smaller than the last. The largest circle, AI, holds the second-largest circle, machine learning. And deep learning is the smallest circle, nested inside of machine learning. Deep learning is the catalyst behind our recently-found, deep-seated AI love interest.

We went over the four kinds of AI in our AI history article. Machine learning (ML) makes our AI possible – its use of algorithms to interpret data is the basis of AI.

With the consistent improvement of computer processing power over the last 50 years, GPUs are faster than ever. And ML algorithms benefit from the highest levels of computer power yet for speed, sizes and amounts of datasets, and broadband connectivity. In ML, machines arrive at conclusions from data that they take in. They can band together and output information for facial, speech, and object recognition, translation, and more. While its basic algorithms are hand-coded, the programmer intends for the machine to start learning by itself when given enough data.

Deep learning (DL) is a branch of machine learning. Its basis lies in artificial neural networks, which are like our own brain’s neural networks, but definitely more hand-crafted. We create layered algorithms that feed information to each other to create the final product or make the final decision. The result is a breakdown of the neural network’s analysis, the system outputs a “probability vector,” which might say something like: “86% confident the object is a human, 55% confident the object is a fruit…”

Not until the DL algorithm is trained over and over again can its probability vector’s confidence grow. After it gets thousands or millions of examples, its probability vector will improve to 98% or 99%. Andrew Ng, who founded the Google Brain project at Google, contributed deeply to the DL field to innovate Google’s image, object, and facial recognition engine. He trained the giant neural network with images from 10 million YouTube videos and created a new frontier for DL.

AlphaGo used DL to fine-tune its neural networks (and thus skills); it played against itself millions of times to train for its match against the human Go champion. But DL isn’t replacing our discerning eye any time soon. A DL that can recognize objects has a difficult time knowing the difference between different species of birds.

AI, ML, and DL are all burgeoning fields that are ripe for innovation. It’s inevitable that we’ll start seeing these technologies more in our everyday lives. Being an AI developer right now isn’t such a bad gig.