Machine Learning and AI with Simple Python and Matlab Scripts (M. Ümit Uyar) (z-library.sk, 1lib.sk, z-lib.sk)

【One-Line Pitch】 A hands-on, math-light courseware that teaches machine learning and AI by having you run and tweak short Python and Matlab scripts on real-world problems—ideal for engineering and science students or professionals who want practical intuition without a heavy theoretical prerequisite. 【Book Arc】 - **Opening (~0%–10%)**: Frames the courseware's audience and goals, then builds the core of neural networks—forward propagation, error/loss, back propagation, weight updates, learning rate, and bias—using a small worked numeric example. - **Early (~10%–30%)**: Turns theory into practice: choosing ANN inputs and outputs, training and saving models, and running complete Python/Matlab scripts for prediction tasks such as a sleep-study grade predictor and a bike-rental demand model. - **Middle (~30%–50%)**: Applies ANNs to domain problems—financial trading with moving-average signals and next-day price prediction, Alzheimer's prognosis from gene-expression data with leave-one-out cross-validation, and NLP-style tweet/keyword inputs for market movement. - **Late (~50%–80%)**: Moves to specialized architectures—convolutional networks for speech recognition and optical character recognition, and recurrent networks for chatbots and language translation—plus typical AI tasks like drone flight control, dietary menu planning, and route planning. - **Ending (~80%–100%)**: Introduces advanced optimization and evolutionary methods (particle swarm optimization, differential evolution, grammatical evolution with a simple genetic algorithm) and closes with remarks on the current state of AI tools; appendices cover bias derivations. 【Book Arc note】Percentages are approximate reading-order markers from the excerpt positions; later chapters are only partially represented in the sample. 【Key Takeaways】 - **Training is error minimization through weight adjustment** (Opening): the book defines loss as the gap between predicted and actual outputs, then shows how gradients and a learning rate update weights iteration by iteration—the conceptual spine for everything later. - **Bias matters in real networks** (Opening): a dedicated section and appendix explain why every layer needs a bias term and how it is trained alongside weights, a detail beginners often skip. - **Input and output selection is a design decision, not a given** (Early): the text stresses that choosing which measurements feed the network—and whether outputs are raw measurements, derived rates, or classifications—shapes whether predictions are useful. - **Complete scripts make the concepts concrete** (Early): end-to-end programs (e.g., a grade predictor that prompts for hours studied/slept, clamps predictions to 0–100, and can save the model as .h5) show the full train-predict-save loop in both Python and Matlab. - **Cross-validation guards against over-fitting** (Middle): the Alzheimer's case uses leave-one-out cross-validation and discusses over-fitting from too many layers/neurons, plus regularization and shuffling as countermeasures. - **Domain context drives architecture choices** (Middle): the trading and NLP examples show how features (moving averages, keyword frequencies, opening price, volume) and network shapes (e.g., 16 inputs, two hidden layers) are chosen with subject-matter input, and why normalization matters for large-valued inputs. - **Architectures specialize by data type** (Late): CNNs for speech and character recognition, RNNs for chatbots and translation—each matched to sequential or spatial structure. - **Beyond gradient descent, evolutionary and swarm methods appear** (Ending): particle swarm optimization, differential evolution, and grammatical evolution (with a simple genetic algorithm) broaden the optimization toolkit beyond back propagation. 【Reading Tips】 - **Deep-read the opening chapters**: the worked forward/back-propagation example is the foundation; work the arithmetic by hand once before running scripts. - **Run every script, then break it**: the exercises repeatedly ask you to change parameters and interpret the impact—treat that as the real learning loop, not optional homework. - **Skim the domain chapters for method, not domain mastery**: you don't need finance or neurology expertise; focus on how features, outputs, and validation are set up. - **Use the Matlab/Python pairing selectively**: pick the language you'll actually use and read the other version only to compare implementation choices. - **Expect non-determinism**: ANN outputs vary between runs; don't chase identical numbers—focus on trends and parameter effects. 【Coverage Limits】 The excerpts cover the ANN fundamentals, several applied cases, and the start of advanced optimization topics, but later chapters (CNN/RNN details, typical AI tasks, evolutionary methods) are only partially represented; this guide should not be treated as a full chapter-by-chapter summary.