Introduction to Machine Learning Concepts
A Conceptual Overview Without Mathematics
1. What Machine Learning Does
Machine learning builds systems that improve their performance on a task by processing data
rather than by following rules written explicitly by a programmer. The rules are inferred from
examples instead of specified in advance.
This approach suits problems where the rules are known implicitly but hard to articulate.
Recognising handwriting, translating language, and detecting fraud are all easier to
demonstrate with examples than to describe procedurally.
It suits other problems poorly. Where requirements are precise, auditable, and stable,
conventional programming remains simpler, cheaper, and easier to verify.
2. Supervised Learning
Supervised learning trains on examples paired with correct answers. The system adjusts
internal parameters to reduce the difference between its outputs and those answers, then is
evaluated on examples it has not seen.
Classification predicts a category, such as whether a message is unwanted. Regression
predicts a continuous quantity, such as a price. The distinction affects which algorithms and
evaluation measures apply.
The dominant practical constraint is labelled data. Obtaining accurate labels at sufficient scale
is frequently the largest cost in a project, and label quality places a ceiling on achievable
performance.
3. Unsupervised and Self-Supervised Learning
Unsupervised learning finds structure in data without labels. Clustering groups similar records;
dimensionality reduction expresses data more compactly while retaining most variation.
Because there is no correct answer to compare against, evaluation is inherently ambiguous.
Cluster results that appear meaningful may reflect arbitrary properties of the algorithm rather
than genuine structure.
Self-supervised learning constructs its own training signal from unlabelled data, for instance by
predicting hidden portions of text or images. This approach underlies most large contemporary
models, since unlabelled data is abundant.
4. Features and Representation
Features are the measurable properties presented to a model. Historically, choosing and
constructing them was the main determinant of performance, and domain knowledge was the
main input.
Deep learning shifted this by learning representations from raw input, reducing manual feature
engineering for images, audio, and text. For structured tabular data, thoughtful feature
construction often still outperforms it.
Feature leakage is a persistent failure mode: a feature that encodes the answer, often through
timing, produces excellent test results and useless production behaviour. Detecting it requires
understanding how data was generated.
5. Training, Validation, and Testing
Data is conventionally divided into a training set for fitting, a validation set for tuning choices,
and a test set reserved for a final estimate of performance. Reusing the test set for tuning
invalidates it.
Cross-validation rotates which portion is held out, producing more stable estimates from
limited data at the cost of additional computation.
Where data has a time dimension, splits must respect chronology. Randomly shuffled splits
allow the model to learn from the future to predict the past, producing estimates that cannot be
reproduced in deployment.
6. Overfitting and Generalisation
Overfitting occurs when a model captures noise specific to its training data rather than the
underlying pattern, performing well on seen examples and poorly on new ones. Underfitting is
the converse: the model is too simple to capture real structure.
Model capacity, data volume, and regularisation govern the balance. Larger models can
memorise more; more data makes memorisation less effective as a strategy; regularisation
penalises complexity directly.
The only reliable evidence of generalisation is performance on data genuinely unseen during
development. Impressive training metrics carry no information about deployment behaviour.
7. Evaluation Measures
Accuracy is inadequate for imbalanced problems. A test for a condition affecting one percent
of a population achieves ninety-nine percent accuracy by always answering negative while
providing no value.
Precision measures how many predicted positives are correct; recall measures how many
actual positives are found. They trade off against each other, and which matters more
depends entirely on the relative cost of the two error types.
Choosing a measure is a decision about consequences, not a technical detail. Optimising a
metric that does not reflect the real objective produces systems that perform well on paper and
badly in use.
8. Neural Networks and Deep Learning
Neural networks compose many simple parameterised operations into layered structures.
Depth allows later layers to build on features detected by earlier ones, producing progressively
more abstract representations.
Their advantage appears with large datasets and unstructured input. With modest tabular
data, simpler methods such as gradient-boosted trees frequently match or exceed them at far
lower computational cost.
Training requires substantial compute and careful configuration. Much practical work involves
adapting existing pre-trained models rather than training from scratch, which reduces both
data and compute requirements considerably.
9. Bias, Fairness, and Interpretability
Models learn patterns present in their training data, including patterns reflecting historical
discrimination. A system trained on past decisions will reproduce the tendencies of those
decisions unless deliberately addressed.
Fairness has multiple formal definitions that cannot generally be satisfied simultaneously.
Choosing among them is a normative decision requiring stakeholder input rather than a
technical optimisation.
Interpretability matters where decisions must be explained or contested. Some methods are
inherently transparent; others require post-hoc explanation techniques whose faithfulness to
the underlying model is itself uncertain.
10. Deployment and Monitoring
A trained model is a small part of a working system. Data pipelines, serving infrastructure,
monitoring, retraining processes, and fallback behaviour typically represent the majority of
engineering effort.
Performance degrades over time as the world changes. Distribution shift, where input data
drifts away from training conditions, is the normal state rather than an exception, making
ongoing monitoring essential.
Feedback loops deserve particular caution. When a model's outputs influence the data it later
trains on, small initial biases can amplify in ways that are difficult to detect from performance
metrics alone.