Machine learning and artificial intelligence are platforms upon which so much of the world is now operated. Companies such as Google, Amazon, Facebook, and Airbnb use them to improve the pace of discovery, improve our perception of what we see on our computer screens, even extract ways to make our smartphones more convenient. Through these and other technologies companies are tapping the human tendency to derive meaning from new experiences and then make them actionable, highly actionable.
I’ve been developing these theories of machine learning for the last few years and a couple of months ago, YouTube invited me to give a talk on them at VidCon. YouTube saw value in having a narrative about how Machine Learning and Artificial Intelligence are being used at the top of YouTube, which is increasingly being used as a gateway to accessing entertainment for young people. So I was asked to come to Southern California to explain it. You can find video footage of my YouTube talk, “The Amazing Ways That YouTube Uses Machine Learning,” below.
A little background. Machine Learning and Artificial Intelligence (ML) are two classes of technologies, with different names. ML is simply a collection of algorithms that incorporate machine learning algorithms (or deep learning), which in turn uses artificial intelligence (AI) models. So artificial intelligence encompasses things like neural networks and deep learning that can be used in computer systems, in fashion (a variety of fashion models that function under multiple environments) and in other fields.
Fundamentally, AI allows computers to make better decisions than humans can based on their mental models, patterns and intuition. They are not perfect but they are better than we are, and the result is a quantum leap in the ways we do things, from how we design and build our computers to how we learn and interact with computers.
In a world where so much of what we see in the digital world is designed by algorithms and underpinned by AI algorithms, YouTube is a fascinating case study. And if we were to study something called “audience reach,” we would be more focused on their success at managing it rather than just on the scaling of the quantity.
YouTube is the most visited online video service in the world.
YouTube’s targeting process is called “Knowledge Graph.” It is based on billions of post-link user reviews and several forms of user behavior, such as views. It generates a knowledge graph with attributes about a specific video, and then uses these attributes to refine the video preferences of its audience, and thereby deliver more compelling and interesting content to those viewers.
For example, YouTube and its parent company Google have developed technology, called Advanced Recommendation System, or Antrix, that uses machine learning to develop a set of tailored opinions for people in your social graph. Most well-known examples are the recommendations for top trending searches which begin with a natural sounding question like, “What’s the hottest band in Denver?” If a video was popular on YouTube, the system will use AI algorithms to also determine if that person is more likely to like it and provide more effective recommendations.
The user experience of AI tools is not flawless, it is also possible that an algorithm might not be better than an actual human being at the task of “knowing” what that person might like. “Googling” is an example of using algorithms to find a matching answer. It is possible that clicking “here” might produce a result we may not like, especially if we encounter it during a period of normal turnover. Indeed, Google is much better than a human at gathering information about the user over time, which is essential for managing (and growing) an audience.
To gather and to understand information about what people like, and therefore what they might be interested in, one needs data about their likes and dislikes. The first step to building this programmatic system is to pick up posts and accounts and identify users.