Machine Learning

A learning problem starts with examples and a task. In supervised learning, examples pair input features with labels that represent the target. Training adjusts a model using examples; inference uses the trained model to produce an output for a new input. Unsupervised learning looks for patterns without supplied target outputs. Self-supervised learning creates surrogate targets from the input data, while transfer learning applies information learned on one task to another.

Choose an approach by asking what information is available and what output is needed. If examples have known answers, prediction may be framed as supervised learning. If the goal is to find structure without answers, an unsupervised approach may fit. These choices describe how learning is organized; they do not by themselves show that the result is useful.

• What is Machine Learning?

  Defines machine learning and describes its learning approaches.

Machine learning

Methods that improve task performance by learning patterns from data or experience.

Training

Adjusting a model using data or experience.

Inference

Using a trained model to produce an output for an input.

Feature

An input property used by a model.

Label

A known answer or target associated with a training example.

Supervised learning

Learning from examples paired with target outputs.

Unsupervised learning

Finding patterns in data without supplied target outputs.

Self-supervised learning

Learning from targets constructed from the input data itself.

Transfer learning

Applying learning from one setting to another task or domain.

Ai Vs Machine Learning

A model is useful when it performs well on inputs beyond the examples used to train it. Overfitting describes a model that fits training examples but performs poorly on new ones. Underfitting describes a model that does not capture useful patterns even in its training data. Generalization is performance on new inputs, and the training data and evaluation design affect how well an evaluation estimates it.

Regularization means adding constraints or penalties intended to reduce unnecessary complexity and overfitting. A hyperparameter is a setting chosen outside the learned parameter values. Use validation data to compare development choices, then reserve test data for a final evaluation. A test set that is repeatedly used to guide model changes can stop providing an independent estimate of performance.

• Datasets: Dividing the original dataset

  Explains how training, validation, and test sets serve different roles in evaluation.

Overfitting

Learning training-data quirks that do not carry over to new data.

Underfitting

Failing to capture useful patterns even in the training data.

Generalization

A model's ability to perform on new inputs beyond its training examples.

Regularization

Constraints or penalties intended to reduce unnecessary model complexity.

Hyperparameter

A setting chosen outside the model's learned parameter values.

A useful evaluation checks performance on examples that were not used to fit the model or guide its development. A validation set can help compare model choices; a test set can provide a final check after those choices are made. Repeatedly consulting either set while changing the model can make its score less informative. Duplicate examples across training and test sets can also make performance look better than it is.

A test set should reflect the data the model is meant to handle, and the metric should match the task. For example, a model that predicts a numeric value and one that assigns a category do different jobs, so the evaluation needs to match the intended output. A valid evaluation is evidence about a defined setting; it does not guarantee performance when the real inputs differ from the evaluation data.

• 3.1. Cross-validation: evaluating estimator performance

  Explains held-out evaluation, cross-validation, overfitting, and risks from tuning against test data.