Deep Learning - a gentle dive by Dries Cronje

Slide 1

Slide 1 text

Deep Learning A gentle dive

Slide 2

Slide 2 text

@dries139 SLACK braai-light

Slide 3

Slide 3 text

Autopilot

Slide 4

Slide 4 text

AlphaGo beats Go world champion AlphaGo beating Sedol

Slide 5

Slide 5 text

Image annotation "black and white dog jumps over bar." “Man in black shirt is playing guitar"

Slide 6

Slide 6 text

Real Time Translation

Slide 7

Slide 7 text

Credit: https://github.com/junyanz/CycleGAN Style Transfer

Slide 8

Slide 8 text

??? "A woman holding a teddy bear in front of a mirror."

Slide 9

Slide 9 text

What is Deep Learning?

Slide 10

Slide 10 text

“deep learning so far has been the ability to map space X to space Y using a continuous geometric transform, given large amounts of human-annotated data. Doing this well is a game-changer for essentially every industry, but it is still a very long way from human-level AI.” François Chollet

Slide 11

Slide 11 text

What is Deep Learning

Slide 12

Slide 12 text

What is Deep Learning X Y Magic Learning phase

Slide 13

Slide 13 text

AlexNet 2012 Feature engineering Deep Learning

Slide 14

Slide 14 text

ImageNet competition 2012 AlexNet 2013 ZFNet 2014 VGG Net 2014 GoogLeNet 2015 Microsoft ResNet And this is how Deep Learning became a big thing

Slide 15

Slide 15 text

Convolutional Neural Network

Slide 16

Slide 16 text

Layered approach

Slide 17

Slide 17 text

Layered approach Layer 1 Layer 2 Layer 3

Slide 18

Slide 18 text

Layered approach a typical layer: 1) Convolutions 2) Activations 3) Pooling Layer 1 Layer 2

Slide 19

Slide 19 text

1. Convolutions Credit: medium.com/towards-data-science

Slide 20

Slide 20 text

1. Convolutions

Slide 21

Slide 21 text

1. Convolutions Edge Detection as an example of a convolution being applied to an image * =

Slide 22

Slide 22 text

1. Convolutions Striding Stride = 1 Credit: adeshpande3.github.io

Slide 23

Slide 23 text

1. Convolutions Striding Credit: adeshpande3.github.io Stride = 1

Slide 24

Slide 24 text

1. Convolutions Padding Credit: https://medium.com/@Synced/ Stride = 1

Slide 25

Slide 25 text

2. Activations ReLU - non-linear activation 1) Replaces sigmoid 2) Less processing power

Slide 26

Slide 26 text

3. Pooling Max Pooling - Reducing dimensionality Credit: cs231n.github.io

Slide 27

Slide 27 text

Choosing Hyperparameters

Slide 28

Slide 28 text

(Convolution + activation + max pooling) + fully connected + softmax Bringing it all together

Slide 29

Slide 29 text