AI Fundamentals
What is Machine Learning?
Add Unite.AI to your preferred sources on Google
Machine learning (ML) is a branch of artificial intelligence in which a system learns patterns from data so it can make predictions, classifications, recommendations, or decisions without a developer writing a separate rule for every possible case. The result is not a machine that “thinks” like a person. It is a statistical model that maps inputs to useful outputs and can be evaluated on data it did not see during training.
Machine learning sits inside the broader field of AI, while deep learning is a family of machine-learning methods built around multi-layer neural networks. This distinction matters: not every AI system uses ML, and not every ML problem requires a neural network.
Key takeaways
- ML learns a relationship from examples rather than relying only on hand-written rules.
- A useful model must generalize to new data, not merely memorize its training set.
- Supervised, unsupervised, semi-supervised, self-supervised, and reinforcement learning solve different kinds of problems.
- Data quality, evaluation design, and monitoring are as important as the algorithm.
How machine learning works
Most ML projects can be understood as a sequence of six stages:
- Define the task. Decide what the system should predict or discover and what success means in the real application.
- Collect and prepare data. Clean records, handle missing values, create useful features, and document where the data came from.
- Split the data. A training set is used to fit the model, a validation set helps select settings, and a test set provides a final estimate on untouched data.
- Train the model. An algorithm adjusts model parameters to reduce a loss function or satisfy another learning objective.
- Evaluate generalization. Metrics must reflect the task, class balance, error costs, and the population on which the model will operate.
- Deploy and monitor. Real-world data can change, so teams watch for drift, performance degradation, bias, and operational failures.
The variables supplied to a model are commonly called features. In supervised learning, the desired answer is called the label or target. A model’s learned parameters encode a relationship between the features and the output; they are not a database of explicit rules.
The main learning paradigms
Supervised learning
In supervised learning, examples include both inputs and known targets. A classification model predicts categories, such as whether a transaction is fraudulent. A regression model predicts a continuous value, such as expected energy demand.
Common supervised algorithms include decision trees, support vector machines, K-nearest neighbors, linear and logistic regression, gradient-boosted trees, and neural networks. A classification threshold such as 0.5 is a decision chosen after a model produces a score or probability; it is not an immutable property of logistic regression.
Unsupervised learning
Unsupervised learning works with data that does not include target labels. The objective may be to find clusters, detect unusual observations, estimate a distribution, or create a lower-dimensional representation. A cluster is a group suggested by a similarity rule; it is not automatically a meaningful real-world class.
Examples include K-means clustering, principal component analysis, density estimation, and some forms of autoencoders. An autoencoder learns to reconstruct its input through a compressed representation. It does not automatically create ground-truth labels.
Semi-supervised and self-supervised learning
Semi-supervised learning combines a small labeled dataset with a larger unlabeled dataset. Self-supervised learning creates a training signal from the data itself—for example, predicting masked words or matching two transformed views of the same image. Self-supervision is central to many modern transformer and foundation-model pipelines because it can use large collections of text, images, audio, or video without a person labeling every example.
Reinforcement learning
In reinforcement learning, an agent takes actions in an environment and receives rewards or costs. The objective is to learn a policy that maximizes expected cumulative reward. This differs from supervised learning because the correct action is not provided for every state, and an action can affect which data the agent encounters next.
Training, validation, and generalization
A model that performs well on its training examples can still fail on new data. This failure is known as overfitting. Teams reduce it through appropriate model capacity, regularization, cross-validation, data augmentation, leakage prevention, and a genuinely independent test set.
There is no single metric for every ML task. Classification may require precision, recall, F1, calibration, or a cost-weighted measure instead of raw accuracy. Regression may use mean absolute error, root mean squared error, or a domain-specific loss. Clustering needs different forms of internal or externally validated evaluation. The metric should reflect what an error means to the user or organization.
Machine learning algorithms are tools, not guarantees
An algorithm carries assumptions. Linear models assume a particular form of relationship. K-nearest neighbors assumes that the selected distance represents meaningful similarity. Naive Bayes assumes features are conditionally independent given the class. Decision trees partition the feature space using learned split rules; their leaves contain predictions based on groups of training observations rather than necessarily one observation each.
Model choice therefore depends on the data size, feature types, latency requirements, interpretability needs, and cost of mistakes. A simpler model can outperform a larger model when data is limited or the operational constraints favor speed and transparency.
Where machine learning is used
ML supports search ranking, recommendations, forecasting, anomaly detection, translation, speech recognition, computer vision, predictive maintenance, fraud detection, and scientific analysis. The same techniques can also amplify historical bias, expose sensitive information, or behave unpredictably under distribution shift. Responsible deployment requires documentation, human oversight where appropriate, security testing, and continued monitoring.
From problem definition to a valid machine-learning experiment
A machine-learning project should begin with a decision and a measurable outcome, not an algorithm. Define the prediction unit, target, observation time, decision time, available features, and cost of each error. For a churn model, for example, using events recorded after cancellation would leak the answer. Establish a simple rule or statistical baseline, then split data by time, customer, location, or another boundary that reflects deployment. Random row splits can place nearly identical observations in training and test sets and produce a misleading score.
Feature engineering converts raw records into representations the model can use, but every feature needs provenance and an availability guarantee. Fit normalization, vocabulary, imputation, and dimensionality reduction only on training data, then apply the learned transformation to validation and test data. Cross-validation estimates variation across samples; a final untouched test set supports the release decision. Select metrics from consequences: precision and recall for unequal classification errors, calibration when probabilities drive action, and cost- or utility-weighted measures when mistakes have different operational effects.
Deployment, monitoring, and responsible operation
Production inference repeats the complete training-time transformation and returns a prediction under latency, throughput, and availability constraints. Package preprocessing with the model, validate input schemas, version artifacts, and compare results between offline and serving implementations. Choose a threshold using the operating capacity and error tradeoff rather than defaulting to 0.5. Roll out through shadow evaluation, a limited cohort, or an experiment with guardrail metrics. Keep a deterministic fallback and a rollback path for dependency failure or unacceptable behavior.
Monitor input quality, feature drift, prediction distribution, calibration, subgroup outcomes, latency, cost, and confirmed labels when they eventually arrive. Drift is a signal to investigate, not automatic proof that retraining will help. Retraining needs reviewed data, repeatable tests, approval, and comparison with the current champion. Document intended and invalid uses, data rights, privacy, security, human override, and appeal where people are affected. Machine learning is a maintained decision system; the model file is only one replaceable component.
Worked example: predicting equipment failure
A manufacturer defines one prediction per machine-day: whether a verified failure will occur within seven days using only telemetry available at the start of that day. It splits by machine and time, compares with age- and threshold-based rules, fits preprocessing on training data, and evaluates event recall, false alerts, warning lead time, calibration, and maintenance capacity. Sensor replacement and planned shutdowns are modeled as operating context rather than treated as ordinary observations.
The model first runs in shadow mode. Alerts show contributing telemetry and uncertainty, but maintainers decide whether to inspect. Findings and confirmed causes become governed labels; absence of a work order is not assumed to mean no failure. A staged rollout uses alert limits and a manual fallback, while monitoring tracks sensor health, input drift, reviewed precision, downtime, and unnecessary maintenance. Retraining occurs only after data and threshold review demonstrates a likely improvement over the current deployed system.
Implementation evidence and operational readiness
A production decision needs more than a successful demonstration. Define the intended users, operating environment, inputs, outputs, dependencies, owner, and the consequence of each important failure. Establish a reproducible baseline and a versioned evaluation set before tuning. Test ordinary cases, boundary conditions, malformed or missing input, distribution shift, dependency outage, misuse, and the groups or environments most likely to be underserved. Measure task quality together with calibration or uncertainty, latency, throughput, resource cost, accessibility, privacy, and security. Record every transformation and threshold so an independent reviewer can reproduce the result and distinguish evidence from an attractive prototype.
Before launch, assign authority for release, exceptions, changes, rollback, and retirement. Use a staged rollout, preserve a safe fallback, and verify monitoring with deliberately injected failures. Operational telemetry should reveal input quality, output behavior, model or rule version, dependency health, human overrides, and confirmed outcomes without collecting unnecessary sensitive data. Define alert thresholds and a response owner, then review real-world evidence after deployment rather than assuming offline performance will persist. Reevaluate whenever data sources, users, models, vendors, policies, hardware, or objectives change. A maintained system also needs documented recovery, incident learning, deletion and retention procedures, and a clear point at which it should be disabled or replaced.
Primary references
