Jahziah.
العربية

STUDY GUIDE · GKU 7

Machine learning study guide (Artificial Intelligence)

This is the heaviest unit of the exam, and it has two halves. The first is classical machine learning: how a model is fitted to data and, more importantly, how you know whether it will work on data it has not seen — training, validation and test sets, cross-validation, the bias–variance trade-off, and the metrics that expose a useless classifier with a high accuracy. Along the way it teaches the standard algorithms by running them by hand: linear and logistic regression with gradient descent, decision trees with entropy and the Gini index, naive Bayes, k-nearest neighbours, support vector machines, k-means and principal components. The second half is neural networks: the neuron and the perceptron rule, activation and loss functions, a full backpropagation step, what makes a network "deep", convolutional and recurrent networks, and how overfitting is detected and controlled. Each section opens with a quick-revision card; the worked examples run from easy to exam level and every number in them can be reproduced with a calculator.

Continue in the app

Structured by the official ETEC subject standard for this topic: every listed topic and learning outcome, with formulas, figures and worked examples. The full sections are available to members in the app.

SKU 7.1 · Machine learning

  1. Learning from data: training, validation and test sets

    What it means for a program to learn from examples, why a model is judged on data it never saw, how the data are split into training, validation and test sets, and how k-fold cross-validation gives a steadier estimate when data are scarce.

  2. Model complexity: bias, variance, overfitting and regularisation

    Why a model that is too simple and a model that is too flexible both fail, the bias–variance decomposition of the expected error, how to read training and test errors to diagnose the two, and how and penalties trade a little bias for a large cut in variance.

  3. Linear and logistic regression trained by gradient descent

    The two workhorse models: linear regression for a numeric target, fitted in closed form or by gradient descent on the mean squared error, and logistic regression for a yes/no target, which passes a linear score through the sigmoid and is fitted by gradient descent on the cross-entropy.

  4. Decision trees and naive Bayes

    Two classifiers that can be built by hand: a decision tree that chooses each split by information gain or Gini impurity, and naive Bayes, which multiplies class priors by per-feature likelihoods, with Laplace smoothing to stop a single zero from deciding the answer.

  5. k-nearest neighbours, support vector machines and choosing an algorithm

    Two geometric classifiers — -NN, which votes among the closest training points, and the support vector machine, which places the boundary with the widest margin — followed by a side-by-side comparison of the algorithms in this unit and how to choose among them.

  6. Supervised versus unsupervised learning: k-means and PCA

    How to tell from the data and the goal whether a task is classification, regression, clustering or reinforcement learning; the k-means algorithm run by hand, the elbow method for choosing , and principal component analysis as the standard way to compress features.

  7. Evaluating classifiers: confusion matrix, precision, recall, F1 and ROC

    The confusion matrix and the metrics built from it, why accuracy misleads on imbalanced data, how the decision threshold trades precision against recall, and how the ROC curve and its area summarise a classifier over all thresholds.

SKU 7.2 · Neural networks and deep learning

  1. Neurons, perceptrons, single-layer and multi-layer networks

    The artificial neuron as a weighted sum passed through an activation, the perceptron learning rule traced to convergence, why one layer cannot learn XOR, how a hidden layer fixes it, and what the universal approximation theorem does and does not promise.

  2. Training a network: activations, loss functions and gradient descent

    The feed-forward pass that turns inputs into a prediction, the common activation functions and their derivatives, the loss functions matched to each kind of output, and how gradient descent, the learning rate, batches and epochs fit together.

  3. Backpropagation in a multi-layer network

    The chain rule organised layer by layer: output delta, hidden deltas, weight gradients, one full update of a 2-2-1 network by hand, and why sigmoid networks suffer vanishing gradients.

  4. Deep learning as a subfield of machine learning

    What "deep" adds to machine learning — representations learned in layers directly from raw data — when it beats classical methods and when it does not, and what it costs in data, computation and interpretability.

  5. Convolutional and recurrent neural networks

    How a convolutional layer slides a small shared kernel over an image, how to compute its output size and parameter count, what pooling does, and how a recurrent network carries a hidden state along a sequence — with the applications of each in vision and language.

  6. Overfitting in neural networks: detection and control

    Reading learning curves to tell overfitting from underfitting and from a run that simply has not converged, and the standard remedies — more or augmented data, early stopping, weight decay, dropout and smaller models — with the arithmetic of each.

MORE ARTIFICIAL INTELLIGENCE GUIDES

All study guides

Read the full guide beside your practice questions.

Open the app