Deep Learning With Very Few Training Examples

Produktbeschreibung

This dissertation addresses the problem of training deep learning models with very few training examples. While deep learning has achieved remarkable success across a wide range of domains, deep learning models typically have a vast number of parameters that need to be optimized, and large amounts of labeled data are required for training. However, the collection and annotation of thousands or millions of examples is extremely time-consuming and expensive. This is a significant limitation of deep learning methods in many real-world applications. In the first part, a novel object detection method is proposed for recognizing new categories with very few training examples by combining the advantages of convolutional neural networks and random forests. Subsequently, a new method called Neural Random Forest Imitation
(NRFI) is presented, designed to implicitly transform random forests into neural networks. The experiments demonstrate that NRFI is scalable to complex classifiers and generates very small networks. Finally, two novel generative methods, ChimeraMix and HydraMix, are presented for small data image classification, which learn the generation of new image compositions by combining the content of multiple images. The methods introduce a feature-mixing architecture for combining multiple images in the feature space guided by a grid-based or segmentation-based mixing mask.

Contents
1 Introduction 1
1.1 Deep Learning . . . . . . . . . . . . . . . . . . . . . . . . . 1
1.2 Small Data . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3
1.3 Contributions . . . . . . . . . . . . . . . . . . . . . . . . . . 5
1.3.1 Part I: Small Data Object Detection . . . . . . . . . 5
1.3.2 Part II: Neural Random Forest Imitation . . . . . . 6
1.3.3 Part III: ChimeraMix & HydraMix . . . . . . . . . . 7
1.4 List of Publications . . . . . . . . . . . . . . . . . . . . . . . 7
1.5 Outline . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 14
2 Related Work 17
2.1 Object Detection . . . . . . . . . . . . . . . . . . . . . . . . 17
2.2 Random Forest to Neural Network Transformation . . . . 18
2.3 Image Classification . . . . . . . . . . . . . . . . . . . . . . 20
2.3.1 Regularization . . . . . . . . . . . . . . . . . . . . . 20
2.3.2 Data Augmentation . . . . . . . . . . . . . . . . . . 20
2.3.3 Mixing Augmentation . . . . . . . . . . . . . . . . . 21
3 Fundamentals 22
3.1 Artificial Intelligence . . . . . . . . . . . . . . . . . . . . . . 22
3.2 Random Forests . . . . . . . . . . . . . . . . . . . . . . . . . 25
3.2.1 Decision Tree . . . . . . . . . . . . . . . . . . . . . . 25
3.2.2 Random Forest . . . . . . . . . . . . . . . . . . . . . 27
3.3 Convolutional Neural Networks . . . . . . . . . . . . . . . 28
3.3.1 Artificial Neuron . . . . . . . . . . . . . . . . . . . . 28
3.3.2 Neural Network . . . . . . . . . . . . . . . . . . . . 30
3.3.3 Training . . . . . . . . . . . . . . . . . . . . . . . . . 32
3.3.4 Convolutional Neural Network . . . . . . . . . . . 34
4 Small Data Object Detection 38
4.1 Method . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 39
4.1.1 Feature Learning . . . . . . . . . . . . . . . . . . . . 40
4.1.2 Random Forest Classification . . . . . . . . . . . . . 40
4.1.3 RF to NN Mapping . . . . . . . . . . . . . . . . . . 40
4.1.4 Fully Convolutional Network . . . . . . . . . . . . 43
4.1.5 Bounding Box Prediction . . . . . . . . . . . . . . . 43
4.2 Localization . . . . . . . . . . . . . . . . . . . . . . . . . . . 44
4.3 Clustering . . . . . . . . . . . . . . . . . . . . . . . . . . . . 46
4.4 Dataset . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 47
4.4.1 Data Capturing . . . . . . . . . . . . . . . . . . . . . 47
4.4.2 Filtering . . . . . . . . . . . . . . . . . . . . . . . . . 48
4.5 Experiments . . . . . . . . . . . . . . . . . . . . . . . . . . . 49
4.5.1 Training and Test Data . . . . . . . . . . . . . . . . 49
4.5.2 Classification . . . . . . . . . . . . . . . . . . . . . . 50
4.5.3 Object Detection . . . . . . . . . . . . . . . . . . . . 51
4.5.4 Computation Time . . . . . . . . . . . . . . . . . . . 53
4.5.5 Localization Accuracy . . . . . . . . . . . . . . . . . 54
4.6 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . 56
5 Neural Random Forest Imitation 57
5.1 Method . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 58
5.1.1 Data Generation . . . . . . . . . . . . . . . . . . . . 58
5.1.2 Imitation Learning . . . . . . . . . . . . . . . . . . . 64
5.2 Experiments . . . . . . . . . . . . . . . . . . . . . . . . . . . 64
5.2.1 Datasets . . . . . . . . . . . . . . . . . . . . . . . . . 65
5.2.2 Implementation Details . . . . . . . . . . . . . . . . 65
5.2.3 Results . . . . . . . . . . . . . . . . . . . . . . . . . . 66
5.2.4 Comparison with State of the Art . . . . . . . . . . 67
5.2.5 Analysis of the Generated Data . . . . . . . . . . . 73
5.3 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . 75
6 ChimeraMix & HydraMix 76
6.1 Method . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 76
6.1.1 Encoder . . . . . . . . . . . . . . . . . . . . . . . . . 77
6.1.2 Mixing Module . . . . . . . . . . . . . . . . . . . . . 78
6.1.3 Decoder . . . . . . . . . . . . . . . . . . . . . . . . . 79
6.1.4 Discriminator . . . . . . . . . . . . . . . . . . . . . . 80
6.1.5 Training . . . . . . . . . . . . . . . . . . . . . . . . . 80
6.2 Experiments . . . . . . . . . . . . . . . . . . . . . . . . . . . 81
6.2.1 Experimental Setup . . . . . . . . . . . . . . . . . . 82
6.2.2 Comparison with State of the Art . . . . . . . . . . 82
6.2.3 Automatic Augmentation . . . . . . . . . . . . . . . 85
6.2.4 CLIP Synset Entropy . . . . . . . . . . . . . . . . . . 88
6.2.5 Generator Impact . . . . . . . . . . . . . . . . . . . 90
6.2.6 Analyses . . . . . . . . . . . . . . . . . . . . . . . . . 90
6.2.7 CLIP Features . . . . . . . . . . . . . . . . . . . . . 94
6.2.8 Hyperparameters . . . . . . . . . . . . . . . . . . . 97
6.3 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . 97
7 Conclusion 98
Bibliography 102

Keywords: Deep Learning, Training mit wenigen Daten, Random Forests, Bildklassifikation, Objekterkennung, Deep Learning, Small Data, Random Forests, Image Classification, Object Detection