SLIDE 1
Want to get started with deep learning? Prototyping a deep learning image classifier
Thomas Ellebæk, Ferring Pharmaceuticals, ML02 PhUSE EU Connect 2018, November 5th 2018
add con
Prototyping a deep learning image classifier Thomas Ellebk, Ferring - - PowerPoint PPT Presentation
Prototyping a deep learning image classifier Thomas Ellebk, Ferring - - PowerPoint PPT Presentation
Want to get started with deep learning? Prototyping a deep learning image classifier Thomas Ellebk, Ferring Pharmaceuticals, ML02 PhUSE EU Connect 2018, November 5th 2018 Pattern detection add con Classification add con {Alligator,
SLIDE 2
SLIDE 3
add con
Classification
{Alligator, Beaver, Cat, Dog, …, Zebra}
SLIDE 4
add con
Classification
{Alligator, Beaver, Cat, Dog, …, Zebra} 0.01 0.00 0.87 0.07 … 0.00
SLIDE 5
add con
Deep learning applications
Teradata report based on survey conducted July 2017:
”80% report that some form of AI is already in production in their organization”
(EB9867_State_of_Artificial_Intelligence_for_the_Enterprises.pdf)
SLIDE 6
add con
Deep learning applications
Teradata report based on survey conducted July 2017:
”80% report that some form of AI is already in production in their organization”
(EB9867_State_of_Artificial_Intelligence_for_the_Enterprises.pdf) www.slideshare.net/AIFrontiers/jeff-dean-trends-and-developments-in-deep-learning-research/8
SLIDE 7
add con
Deep learning breakthrough
ImageNet results (top-5)
Error rate AlexNet 16.4
SLIDE 8
add con
Feedforward Neural Network
SLIDE 9
add con
Feedforward Neural Network
Shoe size Hair length Male Female
SLIDE 10
add con
Feedforward Neural Network
Neuron output: Activation function:
(rectified linear unit – ReLU)
Shoe size and hair length Gender
SLIDE 11
add con
Feedforward Neural Network
Neuron output: Activation function:
(rectified linear unit – ReLU)
Shoe size and hair length Gender
SLIDE 12
add con
Feedforward Neural Network
Neuron output: Activation function:
(rectified linear unit – ReLU)
Shoe size and hair length Gender
SLIDE 13
add con
Feedforward Neural Network
Neuron output: Activation function:
(rectified linear unit – ReLU)
Loss function:
(categorical cross entropy)
Shoe size and hair length Gender
SLIDE 14
add con
Convolutional Neural Network
https://se.mathworks.com/discovery/convolutional-neural-network.html
SLIDE 15
add con
Convolution
2D convolution
a b c d e f g h i j k l m n
- p
Input Kernel
… … … … … … … … e
SLIDE 16
add con
Convolution
2D convolution 2D cross-correlation
a b c d e f g h i j k l m n
- p
Input Kernel
… … … … … … … … e … … … … … … … … e
SLIDE 17
add con
Max pooling
1 4 9 16 2 3 8 15 5 6 7 14 10 11 12 13 4 16 11 14
Input (2,2) max pooling
SLIDE 18
add con
CASE: Cell Counter
SLIDE 19
add con
Computing environment
SLIDE 20
add con
1 2 3 4 >=5 1 79 2 96 3 56 4 157 >=5 52 Predicted labels Actual labels
Fitting AlexNet-type model
Validation accuracy: 35.7%
SLIDE 21
add con
Final model
3 convolutional layers and 2 dense fully connected layers
SLIDE 22
add con
Final model
3 convolutional layers and 2 dense fully connected layers
SLIDE 23
add con
Error analysis
2 correct predicted examples 5 wrongly predicted examples Test accuracy: 55.9% >> 35.7%
SLIDE 24
add con
Live demo!
SLIDE 25
add con
Takeaways and suggested learning
https://machinelearningmastery.com/
SLIDE 26
add con
Takeaways and suggested learning
- Technology is ready
- Data is the most important asset
https://machinelearningmastery.com/
SLIDE 27
add con
Takeaways and suggested learning
- Technology is ready
- Data is the most important asset
https://machinelearningmastery.com/