• Home

  • Documents

  • Lecture 10.3 Introduction to deep learning (CNN) · Introduction to deep learning (CNN) Idar Dyrdal...
23
Lecture 10.3 Introduction to deep learning (CNN) Idar Dyrdal

Lecture 10.3 Introduction to deep learning (CNN) · Introduction to deep learning (CNN) Idar Dyrdal . Deep Learning • Computational models composed of multiple processing layers

  • Upload
    others

  • View
    23

  • Download
    1

Embed Size (px)

Citation preview

Page 1: Lecture 10.3 Introduction to deep learning (CNN) · Introduction to deep learning (CNN) Idar Dyrdal . Deep Learning • Computational models composed of multiple processing layers

Lecture 10.3 Introduction to deep learning (CNN)

Idar Dyrdal

Page 2: Lecture 10.3 Introduction to deep learning (CNN) · Introduction to deep learning (CNN) Idar Dyrdal . Deep Learning • Computational models composed of multiple processing layers

Deep Learning

• Computational models composed of multiple processing layers (non-linear transformations)

• Used to learn representations of data with multiple levels of abstraction: • Learning a hierarchy of feature extractors • Each level in the hierarchy extracts features from the output

of the previous layer (pixels —> classes) • Deep learning has dramatically improved state-of-the-art in:

• Speech and character recognition • Visual object detection and recognition

• Convolutional neural nets for processing of images, video, speech and signals (time series) in general

• Recurrent neural nets for processing of sequential data (speech, text).

2

Level 3

Level 2

Level 1

Raw data (images, video, signals)

Labels

Page 3: Lecture 10.3 Introduction to deep learning (CNN) · Introduction to deep learning (CNN) Idar Dyrdal . Deep Learning • Computational models composed of multiple processing layers

Deep Learning for Object Recognition

«Ship»

Millions of images Millions of parameters Thousands of classes

(AlexNet)

Page 4: Lecture 10.3 Introduction to deep learning (CNN) · Introduction to deep learning (CNN) Idar Dyrdal . Deep Learning • Computational models composed of multiple processing layers

Traditional supervised learning

4

Training images

Feature extraction

Classifier training Classifier

Class labels

Page 5: Lecture 10.3 Introduction to deep learning (CNN) · Introduction to deep learning (CNN) Idar Dyrdal . Deep Learning • Computational models composed of multiple processing layers

Deep learning

5

Training images

Feature extraction

Classifier training Classifier

Class labels

• Learning of weights in the processing layers

• Supervised, unsupervised (or semi-supervised) learning

Page 6: Lecture 10.3 Introduction to deep learning (CNN) · Introduction to deep learning (CNN) Idar Dyrdal . Deep Learning • Computational models composed of multiple processing layers

Semi-supervised learning

6

Labeled samples and (trained) linear decision boundary

Labeled and unlabeled samples and non- linear decision boundary

Page 7: Lecture 10.3 Introduction to deep learning (CNN) · Introduction to deep learning (CNN) Idar Dyrdal . Deep Learning • Computational models composed of multiple processing layers

Artificial Neural Network (ANN) Used in Machine Learning and Pattern Recognition:

• Regression • Classification • Clustering • …

Applications:

• Speech recognition • Recognition of handwritten text • Image classification • …

Network types:

• Feed-forward neural networks • Recurrent neural networks (RNN) • …

7

Feed-forward ANN (non-linear classifier)

Page 8: Lecture 10.3 Introduction to deep learning (CNN) · Introduction to deep learning (CNN) Idar Dyrdal . Deep Learning • Computational models composed of multiple processing layers

Mark 1 Perceptron (Rosenblatt, 1957-59)

8

Cornell Aeronautical Laboratory

Page 9: Lecture 10.3 Introduction to deep learning (CNN) · Introduction to deep learning (CNN) Idar Dyrdal . Deep Learning • Computational models composed of multiple processing layers

Activation functions

Sigmoid (logistic function): Hyperbolic tangent: Rectified linear unit (ReLU):

9

(Quasar Jarosz, English Wikipedia)

Biological neuron:

Page 10: Lecture 10.3 Introduction to deep learning (CNN) · Introduction to deep learning (CNN) Idar Dyrdal . Deep Learning • Computational models composed of multiple processing layers

Feed-forward neural network

10

i

j

k

l

Page 11: Lecture 10.3 Introduction to deep learning (CNN) · Introduction to deep learning (CNN) Idar Dyrdal . Deep Learning • Computational models composed of multiple processing layers

Back-propagation

11

i

j

k

l

Page 12: Lecture 10.3 Introduction to deep learning (CNN) · Introduction to deep learning (CNN) Idar Dyrdal . Deep Learning • Computational models composed of multiple processing layers

Convolutional Neural Network (CNN)

Used in Signal and Image Analysis: • Speech Recognition • Image Recognition • Video Recognition • Image Segmentation • …

Convolutional neural network: • Multi-layer feed-forward ANN • Combinations of convolutional and fully connected layers • Convolutional layers with local connectivity • Shared weights across spatial positions • Local or global pooling layers

12

(A. Karpathy)

Page 13: Lecture 10.3 Introduction to deep learning (CNN) · Introduction to deep learning (CNN) Idar Dyrdal . Deep Learning • Computational models composed of multiple processing layers

Typical CNN

13

(Aphex34)

Learned features

Page 14: Lecture 10.3 Introduction to deep learning (CNN) · Introduction to deep learning (CNN) Idar Dyrdal . Deep Learning • Computational models composed of multiple processing layers

Convolutional neural net

14

Convolution (learned)

Input image

Feature map

Non-linearity

Spatial pooling

Normalisation

Input image

credit: S. Lazebnik

Page 15: Lecture 10.3 Introduction to deep learning (CNN) · Introduction to deep learning (CNN) Idar Dyrdal . Deep Learning • Computational models composed of multiple processing layers

Convolutional neural net

15

Convolution (learned)

Input image

Feature map

Non-linearity

Spatial pooling

Normalisation

Input Feature Map

.

.

.

credit: S. Lazebnik

Page 16: Lecture 10.3 Introduction to deep learning (CNN) · Introduction to deep learning (CNN) Idar Dyrdal . Deep Learning • Computational models composed of multiple processing layers

Convolutional neural net

16

Convolution (learned)

Input image

Feature map

Non-linearity

Spatial pooling

Normalisation Rectified Linear Unit (ReLU)

credit: S. Lazebnik

Page 17: Lecture 10.3 Introduction to deep learning (CNN) · Introduction to deep learning (CNN) Idar Dyrdal . Deep Learning • Computational models composed of multiple processing layers

Convolutional neural net

17

Convolution (learned)

Input image

Feature map

Non-linearity

Spatial pooling

Normalisation Max pooling

credit: S. Lazebnik

Max-pooling: a non-linear down-sampling Provide translation invariance

Page 18: Lecture 10.3 Introduction to deep learning (CNN) · Introduction to deep learning (CNN) Idar Dyrdal . Deep Learning • Computational models composed of multiple processing layers

Convolutional neural net

18

Convolution (learned)

Input image

Feature map

Non-linearity

Spatial pooling

Normalisation

Feature Maps Feature Maps After Contrast Normalization

credit: S. Lazebnik

Page 19: Lecture 10.3 Introduction to deep learning (CNN) · Introduction to deep learning (CNN) Idar Dyrdal . Deep Learning • Computational models composed of multiple processing layers

Convolutional neural net

19

Convolution (learned)

Input image

Feature map

Non-linearity

Spatial pooling

Normalisation

Feature maps after contrast normalization

credit: S. Lazebnik

Page 20: Lecture 10.3 Introduction to deep learning (CNN) · Introduction to deep learning (CNN) Idar Dyrdal . Deep Learning • Computational models composed of multiple processing layers

Example - Caffe Demos

20

Page 21: Lecture 10.3 Introduction to deep learning (CNN) · Introduction to deep learning (CNN) Idar Dyrdal . Deep Learning • Computational models composed of multiple processing layers

21

Page 22: Lecture 10.3 Introduction to deep learning (CNN) · Introduction to deep learning (CNN) Idar Dyrdal . Deep Learning • Computational models composed of multiple processing layers

22

Page 23: Lecture 10.3 Introduction to deep learning (CNN) · Introduction to deep learning (CNN) Idar Dyrdal . Deep Learning • Computational models composed of multiple processing layers

Summary

Topics covered: • Deep learning • Artificial neural networks • Convolutional neural networks.

Further reading: • Szeliski, chapter 14 • Yann LeCun ,Yoshua Bengio & Geoffrey Hinton, “Deep learning”, Nature, Vol 521, 28. May

2015.

Software: • Caffe • MatConvNet (Matlab) • …

23


An Improved Deep Fusion CNN for Image Recognition

An Improved Deep Fusion CNN for Image Recognition

Documents
Learning Deep CNN Denoiser Prior for Image Restorationcslzhang/paper/IRCNN_CVPR17.pdf · Learning Deep CNN Denoiser Prior for Image Restoration Kai Zhang 1; 2, Wangmeng Zuo , Shuhang

Learning Deep CNN Denoiser Prior for Image Restorationcslzhang/paper/IRCNN_CVPR17.pdf · Learning Deep CNN Denoiser Prior for Image Restoration Kai Zhang 1; 2, Wangmeng Zuo , Shuhang

Documents
SMART MANUFACTURING - events.static.linuxfound.org · AGENDA •Introduction •Smart Manufacturing •Applications •Deep learning solution •CNN •Deeplearning4j (DL4J) •Apache

SMART MANUFACTURING - events.static.linuxfound.org · AGENDA •Introduction •Smart Manufacturing •Applications •Deep learning solution •CNN •Deeplearning4j (DL4J) •Apache

Documents
Efficient Hyperparameter Optimization of Deep Learning ... · Optimization of 19 CNN hyperparameters Optimization of 8 CNN hyperparameters Optimization of 15 CNN hyperparameters Optimization

Efficient Hyperparameter Optimization of Deep Learning ... · Optimization of 19 CNN hyperparameters Optimization of 8 CNN hyperparameters Optimization of 15 CNN hyperparameters Optimization

Documents
learning model pit mine, Vietnam using the Mask R-CNN deep

learning model pit mine, Vietnam using the Mask R-CNN deep

Documents
Hybrid Convolutional Recurrent Neural Networks Outperform CNN … · 2020. 2. 3. · study, four deep learning algorithm architectures, including two convention deep learning models

Hybrid Convolutional Recurrent Neural Networks Outperform CNN … · 2020. 2. 3. · study, four deep learning algorithm architectures, including two convention deep learning models

Documents
Convolutional Neural Networks for Deep Learning: An Introduction · 2017-10-05 · Introduction: Convolutional Neural Networks (CNN) Convolutional Neural Networks: A deep learning

Convolutional Neural Networks for Deep Learning: An Introduction · 2017-10-05 · Introduction: Convolutional Neural Networks (CNN) Convolutional Neural Networks: A deep learning

Documents
CNN RNN Autoencoder - UNISTisystems.unist.ac.kr/wp-content/uploads/sites/209/2017/... · 2018. 12. 28. · CNN RNN Autoencoder. Today 1. Machine*Learning 2. Deep*Learning –CNN –RNN

CNN RNN Autoencoder - UNISTisystems.unist.ac.kr/wp-content/uploads/sites/209/2017/... · 2018. 12. 28. · CNN RNN Autoencoder. Today 1. Machine*Learning 2. Deep*Learning –CNN –RNN

Documents
Introduction to Deep Learning - TUM · Introduction to Deep Learning Optimization CNN Introduction to NN Machine Learning basics Back-propagation RNN Prof. Leal-Taixé and Prof. Niessner

Introduction to Deep Learning - TUM · Introduction to Deep Learning Optimization CNN Introduction to NN Machine Learning basics Back-propagation RNN Prof. Leal-Taixé and Prof. Niessner

Documents
Deep Roots: Improving CNN Efficiency With Hierarchical

Deep Roots: Improving CNN Efficiency With Hierarchical

Documents
Deep Label Distribution Learning for Apparent Age Estimation · • CNN for age group classification [Levi & Hassner, CVPR 2015] • DLA based on CNN features from different layers

Deep Label Distribution Learning for Apparent Age Estimation · • CNN for age group classification [Levi & Hassner, CVPR 2015] • DLA based on CNN features from different layers

Documents
Sentiment Analysis Using Deep Learning Technique …acadpubl.eu/jsi/2017-114-7-ICPCIT-2017/articles/11/6.pdf · Sentiment Analysis Using Deep Learning Technique CNN with KMeans

Sentiment Analysis Using Deep Learning Technique …acadpubl.eu/jsi/2017-114-7-ICPCIT-2017/articles/11/6.pdf · Sentiment Analysis Using Deep Learning Technique CNN with KMeans

Documents
Deep-Channel uses deep neural networks to detect single ... · 62 single-molecule analysis17, 18. Convolutional neural network (CNN) layers are a powerful component of 63 deep learning

Deep-Channel uses deep neural networks to detect single ... · 62 single-molecule analysis17, 18. Convolutional neural network (CNN) layers are a powerful component of 63 deep learning

Documents
DEVELOPMENT OF LSTM&CNN BASED HYBRID DEEP LEARNING …

DEVELOPMENT OF LSTM&CNN BASED HYBRID DEEP LEARNING …

Documents
Introduction to Deep Learning · 2020-04-21 · Introduction to Deep Learning. Outline Deep Learning RNN CNN Attention Transformer Pytorch Introduction Basics Examples. RNNs Some

Introduction to Deep Learning · 2020-04-21 · Introduction to Deep Learning. Outline Deep Learning RNN CNN Attention Transformer Pytorch Introduction Basics Examples. RNNs Some

Documents
Convolutional Neural Networks for Biomedical Image ...Deep Learning Intro Perceptron and MLP intro Convolutional NN intro Deep CNN Tools and methods for Deep CNNs. Artificial Neural

Convolutional Neural Networks for Biomedical Image ...Deep Learning Intro Perceptron and MLP intro Convolutional NN intro Deep CNN Tools and methods for Deep CNNs. Artificial Neural

Documents
Point-Voxel CNN for Efficient 3D Deep Learning · Point-Voxel CNN for Efficient 3D Deep Learning Zhijian Liu MIT Haotian Tang Shanghai Jiao Tong University Yujun Lin MIT Song Han

Point-Voxel CNN for Efficient 3D Deep Learning · Point-Voxel CNN for Efficient 3D Deep Learning Zhijian Liu MIT Haotian Tang Shanghai Jiao Tong University Yujun Lin MIT Song Han

Documents
A Short Survey of Deep Learning Methods for 3-D Geometric ......dresser toilet CNN 1 CNN 1 CNN 1 CNN 1 2-D renders from various angles input 3-D shape multi-view CNN output probabilities

A Short Survey of Deep Learning Methods for 3-D Geometric ......dresser toilet CNN 1 CNN 1 CNN 1 CNN 1 2-D renders from various angles input 3-D shape multi-view CNN output probabilities

Documents
Deep Learning for Large-Scale Traffic-Sign Detection and ... · Index Terms—Deep learning, Traffic-sign detection and recog-nition, Traffic-sign dataset, Mask R-CNN, Traffic-sign

Deep Learning for Large-Scale Traffic-Sign Detection and ... · Index Terms—Deep learning, Traffic-sign detection and recog-nition, Traffic-sign dataset, Mask R-CNN, Traffic-sign

Documents
Lecture 6: CNNs and Deep Q Learning 2sudeshna/courses/DL18/CNN-DeepQ-8Mar-2… · Emma Brunskill (CS234 Reinforcement Learning. )Lecture 6: CNNs and Deep Q Learning 49 Winter 2018

Lecture 6: CNNs and Deep Q Learning 2sudeshna/courses/DL18/CNN-DeepQ-8Mar-2… · Emma Brunskill (CS234 Reinforcement Learning. )Lecture 6: CNNs and Deep Q Learning 49 Winter 2018

Documents
Introduction to Deep Learning · 11/19/2018  · November 2018 Introduction to Deep Learning 2. 19. November 2018 Introduction to Deep Learning 3 CNN papers Last time on IDL & open

Introduction to Deep Learning · 11/19/2018  · November 2018 Introduction to Deep Learning 2. 19. November 2018 Introduction to Deep Learning 3 CNN papers Last time on IDL & open

Documents
Image similarity using Deep CNN and Curriculum Learning

Image similarity using Deep CNN and Curriculum Learning

Documents
Active Learning for Deep Object DetectionObject Detection using CNNs An important con-tribution to object detection based on deep learning is R-CNN (Girshick et al., 2014). It delivers

Active Learning for Deep Object DetectionObject Detection using CNNs An important con-tribution to object detection based on deep learning is R-CNN (Girshick et al., 2014). It delivers

Documents
CNN BASICS - Deep Learning

CNN BASICS - Deep Learning

Documents
Toward Accelerating Deep Learning at Scale Using ... · Toward Accelerating Deep Learning at Scale Using Specialized Hardware in the Datacenter . 2 . 3 1Case study: CNN-based Image

Toward Accelerating Deep Learning at Scale Using ... · Toward Accelerating Deep Learning at Scale Using Specialized Hardware in the Datacenter . 2 . 3 1Case study: CNN-based Image

Documents
Evaluation of Deep Learning for Semantic Image ...€¦ · image segmentation in tool condition monitoring. Summary • Deep Learning through CNN can be used for automated semantic

Evaluation of Deep Learning for Semantic Image ...€¦ · image segmentation in tool condition monitoring. Summary • Deep Learning through CNN can be used for automated semantic

Documents
Deep Learning for Emotion Recognition on Small Datasets Using … · 2015. 10. 8. · from movies. We follow a transfer learning approach for deep Con-volutional Neural Network (CNN)

Deep Learning for Emotion Recognition on Small Datasets Using … · 2015. 10. 8. · from movies. We follow a transfer learning approach for deep Con-volutional Neural Network (CNN)

Documents
Deep Learning and Reinforcement Learning Workflows in A.I. · Deep learning based analysis – Various transfer learning methods (Combinations of CNN models, Classifiers) – Estimation

Deep Learning and Reinforcement Learning Workflows in A.I. · Deep learning based analysis – Various transfer learning methods (Combinations of CNN models, Classifiers) – Estimation

Documents
Bankruptcy Prediction by Deep Learningcs230.stanford.edu/projects_winter_2020/reports/32569269.pdf · Deep learning techniques, as the Convolutional Neural Networks (CNN) have only

Bankruptcy Prediction by Deep Learningcs230.stanford.edu/projects_winter_2020/reports/32569269.pdf · Deep learning techniques, as the Convolutional Neural Networks (CNN) have only

Documents
A deeper view of Deep Learning - Telecom Valley...Deep Learning in a nutshell Building a CNN Building an AutoEncoder Tips andricks:T implementation details to build and train oury

A deeper view of Deep Learning - Telecom Valley...Deep Learning in a nutshell Building a CNN Building an AutoEncoder Tips andricks:T implementation details to build and train oury

Documents