Chapter 9 – Making Sense of Machine Learning

9.1  Overview

9.2  The Machine Learning Process

9.2.1  Training Phase

9.2.2  Application Phase

9.2.3  Validation Phase

9.3  Evaluation

9.3.1  Measures of Effectiveness

9.3.2  Precision--Recall Trade-off

9.3.3  Data for Evaluation

9.3.4  Multi-stage Evaluation

9.4  The Fitness Landscape

9.4.1  Hill-Climbing and Gradient Descent/Ascent

9.4.2  Local Maxima and Minima

9.4.3  Plateau and Ridge Effects

9.4.4  Local Structure

9.4.5  Approximating the Landscape

9.4.6  Forms of Fitness Function

9.5  Dealing with Complexity

9.5.1  Degrees of Freedom and Dimension Reduction

9.5.2  Constraints and Dependent Features

9.5.3  Continuity and Learning

9.5.4  Multi-objective Optimisation

9.5.5  Partially Labelled Data

9.6  Summary

accuracy, accuracy measure, application phase, architecture, area under the curve, backpropagation, base rate, centroid, clustering, coherence of clusters, configuration parameters, constraint satisfaction, constraints, continuity in learning, cost–benefit, cross-validation, data reduction, decision tree, deep neural network, degrees of freedom (data), dependent feature, differential (calculus), dimension reduction, ECG , energy landscape, entropy, explainable AI, F score, false negative, false positive, feasible solution, fitness function, fitness landscape, frame of video, frame rate, fully connected, generalisation, genes, genetic algorithm, genetic programming, global search, gradient ascent, gradient descent, hard threshold, hill climbing algorithm, hold out, human labelling, ID3, Industry 4.0, intercept, junk DNA, k-fold cross-validation, k-means algorithm, Kohonen networks, labelling, learning rate, local maximum, local minima/maxima, local minimum, local search, locality, Monte Carlo search, multi-layer neural network, multi-objective optimisation, neural network, neuron, optimal, optimal solution, optimisation, optimisation criterion, optimisation problem, overfitting, Pareto frontier, Pareto, Vilfredo, Pareto-optimal, partially labelled data, perceptron, pinch-point layer, plateau, pre-processing, precision, precision–recall trade-off, principal components analysis, probability, pruning, random forest, recall, reparameterisation, ridge, ROC, semi-supervised learning, sigmoid activation function, sigmoid function, similarity measure, simulated annealing, smoothing, soft constraints, speech recognition, statistical techniques, supervised learning, threshold, training phase, true negative, true positive, unsupervised learning, validation phase, version-space algorithm, visualisation