What Is Machine Learning?

In 1959, Arthur Samuel published a study in the IBM Journal of Research and Development about a program that played checkers. It learned to play better than the person who wrote it, after eight to ten hours of machine-playing time. Samuel wrote: "Programming computers to learn from experience should eventually eliminate the need for much of this detailed programming effort."

Tom Mitchell defined it in his 1997 textbook: a program learns from experience E with respect to a class of tasks T and a performance measure P if its performance at the tasks in T, as measured by P, improves with experience E. His example is checkers again: the task is playing, the measure is the percentage of games won, and the experience is games played against itself.

Machine learning is one part of artificial intelligence, not all of it. Recital 12 of the EU AI Act names two kinds of technique for AI systems: machine learning approaches that learn from data how to achieve certain objectives, and logic- and knowledge-based approaches that infer from encoded knowledge. Deep learning, in turn, is one part of machine learning. According to LeCun, Bengio and Hinton (2015), models made of multiple processing layers learn representations of data at multiple levels of abstraction. An LLM is one such deep model, trained on large volumes of text.

During training, an algorithm adjusts a model's parameters so that its outputs fit the training data. The model is then applied to new data. Google's introduction names four types: supervised, unsupervised and reinforcement learning, plus generative AI, which creates content.

Supervised learning: the model sees many examples where the correct answer is known and discovers the connections that lead to that answer. It can then predict the answer for new cases. One example is emails labelled as spam or not spam, from which a filter learns to sort incoming mail.

Unsupervised learning: the data comes without answers. The model looks for patterns on its own, for instance groups of customers who behave alike. What those groups mean is something the model does not say.

Reinforcement learning: Sutton and Barto describe it as learning which action to take in which situation so as to maximize a numerical reward signal. Nobody tells the learner what to do. It finds out by trying which actions pay off, and the reward often comes later. One example from personalization: in 2010, Yahoo researchers tested a contextual bandit for selecting the articles on the Yahoo front page, a method Sutton and Barto cover in chapter 2 of their textbook. The reward was the click. In the offline evaluation on more than 33 million logged events, the method achieved 12.5% more clicks than a bandit algorithm without user data.

In all three cases the model learns only what is in the data. It inherits the data's errors, gaps and biases.

In marketing, machine learning sits inside tools people use every day. In Smart Bidding, according to Google Ads Help, machine learning algorithms train on data at a vast scale to predict how different bid amounts might affect conversions or conversion value. The system optimizes for the conversions the account reports. Anyone counting the wrong thing there trains the system towards the wrong goal.

Recommendations are the second big field. In 2016, Google researchers described YouTube's as "one of the largest scale and most sophisticated industrial recommendation systems in existence" and reported major performance gains from deep learning.

Search uses it too. Google describes RankBrain as an AI system that helps it understand how words relate to concepts, and neural matching as an AI system that understands representations of concepts in queries and pages and matches them to one another. For SEO, this means Google assesses what a page means, not only which words are on it.

• Check first whether a fixed rule is enough; machine learning pays off where patterns are too complex for hand-written rules.
• Define the task and the performance measure before training a model, like Mitchell's T and P.
• Review training data for gaps, errors and biases before trusting a model with it.
• Test a model on data it did not see during training.
• In Google Ads, report as goals only the conversions that matter to the business.
• Compare results with business figures regularly, because data changes over time.

• Treating machine learning, AI and LLMs as the same thing.
• Assuming a model understands causes when what it finds are relationships in data.
• Training on too little or one-sided data and generalizing the result.
• Optimizing for a metric that does not match the business goal.
• Evaluating a model only on the data it was trained on.
• Changing automated bid strategies constantly and so restarting the learning again and again.