Appunti di Machine Learning

Appunti di Machine Learning

Sommario

Capitolo 1 - Introduzione al Machine Learning ................................................................................................. 6

Task .................................................................................................................................................................... 6

Performance ...................................................................................................................................................... 6

Tipi di learning ................................................................................................................................................... 7

Supervised Learning ...................................................................................................................................... 7

Semi-supervised Learning

.............................................................................................................................. 7

Unsupervised Learning .................................................................................................................................. 7

Reinforcement Learning ................................................................................................................................ 7

Training data ...................................................................................................................................................... 8

Tipi di dati .......................................................................................................................................................... 8

Parametri e Hyper-parametri ............................................................................................................................ 9

Overfitting e Underfitting .................................................................................................................................. 9

Linear Regression......................................................................................................................................... 10

Complessità del modello ............................................................................................................................. 11

No free lunch Theorem ................................................................................................................................... 12

Bias and Variance error ................................................................................................................................... 12

Generalization vs Model Capacity ................................................................................................................... 13

Test Set ............................................................................................................................................................ 14

Validation Set .................................................................................................................................................. 14

k-Fold cross validation ..................................................................................................................................... 14

Leave-one-out validation ................................................................................................................................. 15

Dataset Augmentation .................................................................................................................................... 15

Regularization .................................................................................................................................................. 16

Tikhonov regularization ............................................................................................................................... 17

Performance evaluation .................................................................................................................................. 17

Accuracy ...................................................................................................................................................... 17

Classification or Confusion Matrix ............................................................................................................... 18

Precision and Recall ..................................................................................................................................... 18

F-Score/ F-Sndex .......................................................................................................................................... 19

Tunable System ........................................................................................................................................... 19

ROC Curve .................................................................................................................................................... 19

Capitolo 2 - Neural Networks .......................................................................................................................... 20

McCulloch & Pitts neuron ................................................................................................................................ 20

Rosenblatt’s Perceptron .................................................................................................................................. 21

Feed-forward networks ................................................................................................................................... 22

Connections ................................................................................................................................................. 23

Multi-Layer Perceptron ................................................................................................................................... 24

Universal Approximation Theorem ............................................................................................................. 25

Gradient Descent ............................................................................................................................................. 26

Back Propagation ............................................................................................................................................. 27

Stochastic Gradient Descent ........................................................................................................................... 29

Early Stopping .............................................................................................................................................. 30

Momentum .................................................................................................................................................. 31

Adaptive Learning Rate ................................................................................................................................... 32

Regularization .................................................................................................................................................. 32

Activation Functions ........................................................................................................................................ 32

Regression problem ..................................................................................................................................... 33

MLP as binary classifier ................................................................................................................................... 33

MLP as multi-class classifier ............................................................................................................................ 34

Capitolo 3 - Competitive Neural Networks...................................................................................................... 35

Learning Vector Quantization.......................................................................................................................... 35

Kohonen neuron .............................................................................................................................................. 35

Unsupervised LVQ ........................................................................................................................................... 36

Training Phase ............................................................................................................................................. 36

Supervised LVQ ................................................................................................................................................ 37

Training Phase ............................................................................................................................................. 37

Neuron under-utilization ................................................................................................................................. 38

Solution ........................................................................................................................................................ 38

Comparison with MLP ..................................................................................................................................... 39

Manifolds learning ........................................................................................................................................... 39

Self-Organizing Maps (SOM) ........................................................................................................................... 40

Learning algorithm ...................................................................................................................................... 41

Manifold 1D ................................................................................................................................................. 41

Manifold 2D ................................................................................................................................................. 42

Capitolo 4 – Deep learning .............................................................................................................................. 44

Representation learning .................................................................................................................................. 45

Transfer Learning ............................................................................................................................................. 47

Convolutional neural networks (CNN) ............................................................................................................. 48

Convolution ................................................................................................................................................. 48

Convolutional layers .................................................................................................................................... 49

Layer organization ....................................................................................................................................... 50

Receptive Field ............................................................................................................................................ 50

Stride ........................................................................................................................................................... 50

Padding ........................................................................................................................................................ 52

Pooling layer ................................................................................................................................................ 53

Dropout layer............................................................................................................................................... 53

Output layers ............................................................................................................................................... 54

AlexNet ........................................................................................................................................................ 54

Advanced Keras Models .................................................................................................................................. 55

Computational Graph .................................................................................................................................. 56

Keras functional API

..................................................................................................................................... 57

Input Layer ................................................................................................................................................... 57

Tensor operations ........................................................................................................................................ 58

Creating Tensors with network layers ......................................................................................................... 59

Non-sequential models ............................................................................................................................... 60

Weight sharing............................................................................................................................................. 61

Custom loss functions .................................................................................................................................. 61

Generating new samples during the training .............................................................................................. 63

Data augmentation on images .................................................................................................................... 63

Fine tuning ................................................................................................................................................... 65

Learning Strategies for Deep Networks........................................................................................................... 65

Skip Connections ......................................................................................................................................... 67

Batch Normalization .................................................................................................................................... 68

Greedy Supervised Pre-Training .................................................................................................................. 70

Auxiliary Heads ............................................................................................................................................ 70

Depthwise Separable Convolutional layers ..................................................................................................... 71

Transposed Convolutional Layers ................................................................................................................ 73

Autoencoders .............................................................................................................................................. 74

Variational AutoEncoders (VAE) .................................................................................................................. 75

Cap 5 – Reinforcement Learning ..................................................................................................................... 78

Episodes ....................................................................................................................................................... 78

Policy............................................................................................................................................................ 80

Q-learning .................................................................................................................................................... 81

Practical aspects .......................................................................................................................................... 83

Replay Buffer ............................................................................................................................................... 85

Generative Adversarial Networks.................................................................................................................... 85

Adversarial Machine Learning ......................................................................................................................... 87

GAN .................................................................................................................................................................. 87

Training of GAN ........................................................................................................................................... 88

The loss function.......................................................................................................................................... 89

Training algorithm ....................................................................................................................................... 90

Mode Collapse ............................................................................................................................................. 91

Supervision with labels solution .................................................................................................................. 91

Reward sample diversity solution ............................................................................................................... 91

Example ....................................................................................................................................................... 92

Recurrent Networks ........................................................................................................................................ 93

Loss Function ............................................................................................................................................... 95

Task 1: sequence to sequence ..................................................................................................................... 95

Task 2: sequence to single value ................................................................................................................. 95

Task 3: single value to sequence ................................................................................................................. 96

Task 4: sequence to sequence with different lengths ................................................................................. 96

The memory problem ....................................................................................