This morning, machine learning has probably already done its job for you three times before you even opened your first email. It calculated your route to work, filtered spam out of your inbox, and suggested a video for you. Still, for most people, it\u2019s unclear what\u2019s happening behind the scenes. This article explains how machine learning works\u2014without any formulas\u2014from raw data through training to the final prediction. And at the end, it shows you what you really need to get started on your own.<\/p>\n\n\n\n
Machine learning is a process in which a computer learns from examples rather than following fixed rules. It is shown large amounts of historical data, independently identifies patterns within it, and applies these patterns to new, unknown cases. The result is not a definitive answer, but a probability. Humans do not program the solution; they provide the data.<\/p>\n\n\n\n
To understand what makes this special, it helps to look at traditional software. A typical program follows rules that a human has written down in advance: If case A occurs, do B. This works well as long as all cases can be specified in advance.<\/p>\n\n\n\n
But that doesn't work for many tasks. No one can define, using fixed rules, how to recognize a cat in a photo or when an email is spam. There are too many exceptions. This is exactly where machine learning turns the logic on its head. Instead of giving the computer the rules, you give it examples and the correct answers and let it figure out the rules on its own.<\/p>\n\n\n\n
Here\u2019s an example: To build a traditional spam filter, you\u2019d have to manually enter every suspicious phrase. With machine learning, however, you instead show the system tens of thousands of emails that have already been marked as \u201espam\u201c or \u201enot spam.\u201c From this, the model deduces on its own which characteristics are typical of spam\u2014including those that no human would have thought of.<\/p>\n\n\n\n
A machine learning model isn't created in a single step, but rather through an iterative process. In practice, data teams break it down into five phases.<\/p>\n\n\n\n
Collect data.<\/strong> It all starts with the question to be answered and the data needed to answer it. This could be sales figures from your own system, sensor data from a machine, or publicly available datasets. Without a sufficiently large, representative dataset, the rest won\u2019t work.<\/p>\n\n\n\n Prepare the data.<\/strong> Raw data is almost always incomplete, incorrect, or duplicated. It must be cleaned and standardized. This is followed by the step that requires the most work in traditional machine learning: the so-called Feature Engineering<\/a>, in which experts analyze the data to identify the features that the model can use to recognize patterns. This preparatory work traditionally takes up the majority of the time in data projects.<\/p>\n\n\n\n Train the model.<\/strong> The algorithm now receives the processed data. It searches for patterns within the data and continuously adjusts its internal values to minimize its errors. This process is repeated many times until the predictions are reliable enough.<\/p>\n\n\n\n Check the model.<\/strong> Then comes the moment of truth. The model is tested with data it has never seen during training. This reveals whether it has truly learned or is merely memorizing. The most common mistake is called overfitting: The model has \u201ememorized\u201c the training data, including its random variations, and fails as soon as the real data deviates even slightly.<\/p>\n\n\n\n Insert and observe.<\/strong> If the model passes the test, it goes into production and makes decisions based on new data. It\u2019s never really finished. The world changes, and so does the data. That\u2019s why a model must be monitored and regularly retrained with fresh data; otherwise, its performance will deteriorate over time.<\/p>\n\n\n\n \u201eMachine learning\u201c is an umbrella term for several approaches. Which one is appropriate depends on the available data and the objective.<\/p>\n\n\n\n Supervised Learning<\/strong> This is the most common scenario in practice. The model learns from examples where the correct answer is already known, such as emails labeled \u201espam\u201c or \u201enot spam.\u201c If the answer is a category, this is called classification. If it is a number\u2014such as the expected revenue for next month\u2014it is called regression.<\/p>\n\n\n\n Unsupervised Learning<\/strong> works with data without predefined answers. The model\u2019s task is to identify patterns on its own\u2014for example, customer groups with similar behavior. Because there is no fixed goal, experts must interpret the results afterward.<\/p>\n\n\n\n Reinforcing Learning<\/strong> does not learn from a fixed dataset, but rather through trial and error. A system receives a reward for good decisions and a penalty for bad ones, and develops its own strategy over many iterations. This approach is behind many breakthroughs in robotics.<\/p>\n\n\n\n Deep Learning<\/strong> is a special case that involves artificial neural networks<\/a> and loosely mimics human cognition. Its advantage: It can directly process unstructured data such as images, speech, or sound and identifies the relevant features on its own. The trade-off is that it requires very large amounts of data and a great deal of computing power.<\/p>\n\n\n\n How Deep learning is distinct from machine learning and artificial intelligence<\/a>, is a separate question.<\/p>\n\n\n\nThe Four Types of Machine Learning<\/h2>\n\n\n\n