Convolutional Neural Networks

So Far - Fully Connected Networks

Flattened Input

Full Connectivity

Courtesy: Keelin Murphy

@MurphyKeelin

Image Analysis

256

256

Fully Connected Model:

W1 will have size 256 x 256 x 3

W1, b1

flatten

Loss of spatial information

Input Layer

First Hidden Layer

Wm, bm

Courtesy: Keelin Murphy

@MurphyKeelin

Convolutional Neural Networks

Input Layer

First Hidden Layer

Feature 1

Feature 2

Feature 3

Feature 4

Feature 5

1) Each "channel" of neurons represents 1 feature

2) Each neuron connects only to a small "receptive field"

3) Neurons in the same channel share the same weights (detect the same feature in different locations)

Channels

Courtesy: Keelin Murphy

@MurphyKeelin

Convolutional Neural Networks

Output after applying weights to input data

Courtesy: Keelin Murphy

@MurphyKeelin

Learning Features in Layers

Early Layers

Final Layers

Courtesy: Keelin Murphy

@MurphyKeelin

Terminology

Padding:

Patch-size=

Filter-size=

Kernel-size=

3x3

Stride=1x1

Padding=valid

Padding=same

Courtesy: Keelin Murphy

@MurphyKeelin

Network Layer Types

Convolutional Layer

Fully-Connected 

(Densely-Connected) Layer

Max-Pool Layer

Dropout

Layer

Courtesy: Keelin Murphy

@MurphyKeelin

Tensorflow Convolutional Tutorial

Softmax

Courtesy: Keelin Murphy

@MurphyKeelin

Resources for Deep-Learning

Courses:

https://www.coursera.org/specializations/deep-learning

http://www.fast.ai

 

General:

http://cs231n.github.io/

http://neuralnetworksanddeeplearning.com

http://deeplearning.stanford.edu/wiki/index.php/UFLDL_Tutorial

https://ayearofai.com

https://machinelearningmastery.com/

 

Research:

https://arxiv.org/  (for research literature)

 

Other:

https://github.com/kailashahirwar/cheatsheets-ai (cheat-sheets for programming)

https://aiexperiments.withgoogle.com (fun stuff with neural nets)

Courtesy: Keelin Murphy

@MurphyKeelin

Made with Slides.com