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 Ellebæk, Ferring Pharmaceuticals, ML02 PhUSE EU Connect 2018, November 5th 2018

Pattern detection add con

Classification add con {Alligator, Beaver, Cat , Dog, …, Zebra}

Classification add con {Alligator, Beaver, Cat , Dog, …, Zebra} 0.01 0.00 0.87 0.07 … 0.00

Deep learning applications Teradata report based on survey conducted July 2017: add ”80% report that some form of AI is already in production in their organization ” con (EB9867_State_of_Artificial_Intelligence_for_the_Enterprises.pdf)

Deep learning applications Teradata report based on survey conducted July 2017: add ”80% report that some form of AI is already in production in their organization ” con (EB9867_State_of_Artificial_Intelligence_for_the_Enterprises.pdf) www.slideshare.net/AIFrontiers/jeff-dean-trends-and-developments-in-deep-learning-research/8

Deep learning breakthrough ImageNet results (top-5) add con 16.4 Error rate AlexNet

Feedforward Neural Network add con

Feedforward Neural Network add con Shoe size Male Hair length Female

Feedforward Neural Network add Neuron output: con Activation function: (rectified linear unit – ReLU) Shoe size and Gender hair length

Feedforward Neural Network add Neuron output: con Activation function: (rectified linear unit – ReLU) Shoe size and Gender hair length

Feedforward Neural Network add Neuron output: con Activation function: (rectified linear unit – ReLU) Shoe size and Gender hair length

Feedforward Neural Network add Neuron output: con Activation function: (rectified linear unit – ReLU) Loss function: (categorical cross entropy) Shoe size and Gender hair length

Convolutional Neural Network add con https://se.mathworks.com/discovery/convolutional-neural-network.html

Convolution add Kernel con 2D convolution Input a b c d … … e e f g h i j k l m n o p … … … … … …

Convolution add Kernel con 2D convolution 2D cross-correlation Input a b c d … … … … e e e f g h i j k l m n o p … … … … … … … … … … … …

Max pooling add con Input 1 4 9 16 (2,2) max pooling 2 3 8 15 4 16 11 14 5 6 7 14 10 11 12 13

CASE: Cell Counter add con

Computing environment add con

Fitting AlexNet-type model add con Predicted labels 1 2 3 4 >=5 Validation accuracy: 35.7% 1 0 0 0 79 0 Actual labels 2 0 0 0 96 0 3 0 0 0 56 0 4 0 0 0 157 0 >=5 0 0 0 52 0

Final model add con 3 convolutional layers and 2 dense fully connected layers

Final model add con 3 convolutional layers and 2 dense fully connected layers

Error analysis 2 correct predicted examples Test accuracy: 55.9% >> 35.7% add con 5 wrongly predicted examples

Live demo! add con

Takeaways and suggested learning add con https://machinelearningmastery.com/

Takeaways and suggested learning add con ‣ Technology is ready ‣ Data is the most important asset https://machinelearningmastery.com/

Takeaways and suggested learning add con ‣ Technology is ready ‣ Data is the most important asset Questions? https://machinelearningmastery.com/