What is Machine Learning?

Machine learning (ML) is a subfield of artificial intelligence in which computer systems learn from data instead of being explicitly programmed for every task. A learning algorithm analyzes sample data, identifies patterns and relationships within it, and builds a statistical model from them. That model can then be applied to new, previously unseen data – to make predictions, classify objects or derive decisions.

The decisive difference from conventional software: rule-based programs and expert systems follow rigid if-then instructions that people defined in advance. A machine learning system, by contrast, derives its rules independently from the training data – and becomes continuously more accurate with more and better data. That is precisely what makes machine learning the key technique behind modern artificial intelligence: from spam filters and product recommendations in online shops to the large language models behind applications such as ChatGPT.

To top

Every ML project starts with data. During the training phase, the learning algorithm processes a training dataset – thousands of images, sensor readings or texts, for example – and gradually adjusts the internal parameters of its model so that the deviation between prediction and actual result keeps shrinking. In other words, the model learns which features in the inputs relate to which output values: which pixel patterns indicate a defective component? Which transaction characteristics point to fraud?

Whether what has been learned holds up only becomes clear with data the model has never seen. The dataset is therefore usually split: the largest portion is used for training, the remainder for testing predictive accuracy. Only once the test results are convincing does the model go into productive use – known as inference, in which it continuously evaluates new data points and draws conclusions from them.

The quality of an ML model depends less on the algorithm than on the data: if the training data is incomplete, outdated or one-sided, the model inherits those weaknesses. “Garbage in, garbage out” applies without exception in machine learning, which is why data preparation and feature engineering often take up more time in practice than the actual training.

Three types of machine learning are traditionally distinguished – according to the kind of feedback the learning algorithm receives during training.

Supervised learning: The model learns from labeled sample data in which the correct answer is already known – emails flagged as “spam” or “not spam”, for instance. From these pairs of input and desired output, the system learns the relationship and can transfer it to new cases. Typical tasks are classification (assignment to categories) and regression (predicting numerical values such as prices or utilization). Supervised learning is the most widely used learning method in practice.

Unsupervised learning: Here the algorithm receives unlabeled data and has to discover structures in it on its own. Methods such as clustering group similar data points – customers with comparable buying behavior, for example – without the groups having been defined beforehand. Anomaly detection, such as for unusual network access or machine readings, is also mostly unsupervised.

Reinforcement learning: An agent learns by trial and error within an environment: it receives a reward for actions that lead to the goal and a penalty for mistakes. Over many iterations it develops a strategy that maximizes the reward. The method became widely known through systems such as AlphaGo; today it is used in robotics, in the control of autonomous vehicles and in fine-tuning large language models, among other areas.

Semi-supervised learning is occasionally named as a fourth type: it combines a small amount of labeled data with large volumes of unlabeled data – useful when labeling by specialists is expensive, as with medical imaging.

To top

The three terms relate to one another like concentric circles. Artificial intelligence is the umbrella term for all systems that solve tasks normally assumed to require human intelligence. Machine learning is the most important subfield of AI – the data-driven approach through which most modern AI systems acquire their behavior. Deep learning, in turn, is a subset of machine learning based on artificial neural networks with many processing layers.

The practical difference lies in how features are handled: in classical machine learning, specialists largely define the relevant features of the data themselves, and the algorithm “only” learns how they relate to the outcome. Deep learning models, by contrast, extract features automatically from raw data such as images, audio or text – but they require considerably larger volumes of data and many times the computing power.

Generative AI systems such as ChatGPT are applied machine learning as well: large language models are trained on vast amounts of text using deep learning and then refined with reinforcement learning. The chain of terms thus closes in on itself – LLMs are deep learning, deep learning is machine learning, and machine learning is artificial intelligence.

Different tasks have given rise to different ML algorithms. The most important ones include:

Decision trees and random forests: Decision trees break a decision down into a sequence of simple yes/no questions, which makes them easy to follow. A random forest combines many such trees into an ensemble that is considerably more robust than a single tree. A related approach is gradient boosting, which trains trees one after another, each correcting the errors of its predecessors – often the most accurate method for structured data.

Linear and logistic regression: These statistical classics predict numerical values and estimate probabilities for categories respectively. They are fast, economical and easy to interpret – and therefore still in productive use despite all the advances since, for example in risk assessment.

Support vector machines (SVM): SVMs separate data classes by a boundary with the greatest possible distance to the nearest data points. They deliver precise classifiers even with smaller datasets and were the standard in pattern recognition before the rise of deep learning.

k-means clustering: The best-known algorithm in unsupervised learning divides data points into a predefined number of groups – the basis for many customer segmentations and anomaly analyses.

Artificial neural networks: Inspired by the workings of the brain, they process inputs through layers of connected computing units whose weights are adjusted during training. With many layers they become deep neural networks – the foundation of deep learning and therefore of image, speech and text processing at today’s level.

In industry, machine learning has long been part of everyday operations. Predictive maintenance uses sensor data to forecast when a machine will need servicing, before it fails. Camera-based quality control detects manufacturing defects more reliably than the human eye, and in logistics ML models optimize routes, inventory levels and demand forecasts.

Machine learning is part of daily life for everyone, too: spam filters classify emails, streaming services and online shops calculate recommendations, voice assistants convert speech into text, navigation systems predict traffic jams, and banks detect fraudulent card payments in real time. In medicine, ML models support image analysis in radiology and the search for new active substances.

Machine learning holds particular importance in science: climate research, materials development, genome analysis and particle physics generate volumes of data that could no longer be evaluated without learning systems. Models like these are trained on supercomputers – where machine learning and classical simulation merge into a new research instrument.

The biggest hurdle in machine learning is rarely the algorithms, but the data: it has to be available in sufficient quantity, quality and representativeness. Biased training data leads to biased models – a problem with real consequences from lending to recruitment, and one that has to be addressed through careful data selection and validation.

A second core problem is overfitting: an overfitted model effectively memorizes the training data – including its random quirks – and then fails on new data. The counterpart, underfitting, describes a model that is too simple to capture the relevant relationships in the first place. Techniques such as cross-validation, regularization and a strict separation of training and test data keep both effects in check.

Then there is the question of interpretability: the more complex a model, the harder it is to trace why it arrives at a particular decision. In regulated areas – medicine, finance or under the EU AI Act, for example – explainability is mandatory. And finally, the demand for computing power grows with model size: modern ML methods, deep learning above all, can no longer be trained sensibly without specialized hardware.

Training modern ML models is at heart a problem of high performance computing: millions to billions of parameters are adjusted across countless iterations on enormous datasets. These massively parallel matrix computations are the domain of the GPU – and, as soon as a model no longer fits on a single server, the domain of GPU clusters with fast interconnect, a parallel file system and efficient cooling.

What ML infrastructure looks like in practice is shown by the research systems built by MEGWARE: the GPU cluster Helma at the University of Erlangen, used among other things to train AI models, and Capella at TU Dresden, one of the most energy-efficient systems on the Green500 list. The direct hot-water cooling of the EUREKA platform also makes the power density of modern GPU generations economically manageable – and the waste heat can even be used to heat buildings.

For companies and research institutions looking to put machine learning into production, the infrastructure question is therefore a strategic one: cloud, in-house cluster or hybrid operation – what matters is utilization, data protection and total cost. MEGWARE supports these projects from requirements analysis through system architecture to operation; in the Benchmark Center, you can test your own ML workloads on real hardware before investing. For an overview, see MEGWARE’s solutions for artificial intelligence.

What is machine learning explained simply?

Machine learning means a computer learns from examples rather than from hard-coded rules. It is given a large amount of data, identifies patterns in it and then applies what it has learned to new cases – much like a spam filter learns to recognize unwanted emails by their typical characteristics.

What are the three types of machine learning?

Supervised learning (learning from labeled examples), unsupervised learning (independently discovering structures in unlabeled data) and reinforcement learning (learning through reward and penalty). Semi-supervised learning, which combines both kinds of data, is often added as a fourth type.

What is the difference between AI and machine learning?

Artificial intelligence is the umbrella term for all systems with “intelligent” behavior – including rule-based programs. Machine learning is the subfield of AI in which systems learn their behavior from data. Virtually every modern AI application today is based on machine learning.

What is the difference between machine learning and deep learning?

Deep learning is a subcategory of machine learning that uses deep neural networks with many layers. It extracts features automatically from raw data and is particularly strong with images, speech and text – but it requires considerably more data and computing power than classical ML methods.

Is ChatGPT machine learning?

Yes. ChatGPT is based on a large language model (LLM) that was trained on vast amounts of text using deep learning and then refined with reinforcement learning from human feedback – a combination of several machine learning methods.

What is overfitting?

Overfitting means a model learns the training data too precisely – including noise and random patterns – and therefore delivers poor predictions on new data. The remedies are more or more varied training data, regularization and validation against separate test data.

What are examples of machine learning in everyday life?

Spam filters, product and film recommendations, voice assistants, face recognition on smartphones, traffic forecasts in navigation apps, fraud detection for card payments and the automatic sorting of photos – trained ML models sit behind all of them.

What hardware does machine learning need?

Classical ML methods such as decision trees run on ordinary CPU servers. Deep learning, on the other hand, requires GPUs with specialized computing units; large models are trained on GPU clusters with high-speed interconnect – in other words on HPC infrastructure of the kind MEGWARE develops and builds for research and industry.