ݺߣ

ݺߣShare a Scribd company logo

Zenodia TechDays talks Oct 24-25 Stockholm Kistamässan

•
•
Zenodia Charpy
Zenodia Charpy

TechDays Stockholm presentation understand how AI learns train an AI model to identify Pneumonia deploy an AI model on Azure as a web service with databrick comsume the model generalize the model to other scenario ( flappy bird, CT scan segmentation...etc)

1 of 14
1. INTUITION – HOW AI LEARN 2. PUT IT IN CONTEXT – PNEUMONIA 3. PLAN OF ATTACK– HOW TO DO IT 4. DEMO – JUPYTER NOTEBOOK 5. GENERALIZATION
Is this an apple ? apple apple not apple apple apple not appleapple apple not apple ...
apple apple ... not appleapple apple not apple Neural Network Training Validataion Labels Features apple apple not apple 1/3 2/3Performance : (accuracy) 2/3 =67% Prediction apple apple apple
INPUTs : (image + label) Feed to NN NN predicts labels BUILD A MODEL TO CLASSIFY PNEUMONIA The Neural Networks Model Prediction . . Pneumonia Normal Normal Pneumonia ...
All done in one Jupyter notebook 1.Get data 2.Data Process & Explore 3.1 framework (TF, Keras, Torch...etc) 3.2 loss & optimize function 3.Neural Network architecture Deep learning 4.Prediction 3.3 train & validate
DEMO Demo
Zenodia TechDays talks Oct 24-25 Stockholm Kistamässan
THE SAME CONCEPT CAN BE APPLIED TO OTHER SCENARIOS
INPUTs : Images of objects of interests Feed to NN Object shapes BUILD A MODEL TO LEARN VIDEO SEMANTIC SEGMENTATION (ENET) Model Prediction: The Neural Networks
INPUTs : Images of objects of interests Feed to NN Object labels + location BUILD A MODEL TO LEARN OBJECT DETECTION IN VIDEO (SSD) Model Prediction: The Neural Networks
Raw input abdominal CT scan Output segment BUILD A MODEL TO SEGMENT 4D CT SCAN ORGANS (NIFTYNET)
BUILD A MODEL TO LEARN TO PLAY ANGRY BIRD
PARKING
Predictedoriginal

More Related Content

Zenodia TechDays talks Oct 24-25 Stockholm Kistamässan

  • 1. 1. INTUITION – HOW AI LEARN 2. PUT IT IN CONTEXT – PNEUMONIA 3. PLAN OF ATTACK– HOW TO DO IT 4. DEMO – JUPYTER NOTEBOOK 5. GENERALIZATION
  • 2. Is this an apple ? apple apple not apple apple apple not appleapple apple not apple ...
  • 3. apple apple ... not appleapple apple not apple Neural Network Training Validataion Labels Features apple apple not apple 1/3 2/3Performance : (accuracy) 2/3 =67% Prediction apple apple apple
  • 4. INPUTs : (image + label) Feed to NN NN predicts labels BUILD A MODEL TO CLASSIFY PNEUMONIA The Neural Networks Model Prediction . . Pneumonia Normal Normal Pneumonia ...
  • 5. All done in one Jupyter notebook 1.Get data 2.Data Process & Explore 3.1 framework (TF, Keras, Torch...etc) 3.2 loss & optimize function 3.Neural Network architecture Deep learning 4.Prediction 3.3 train & validate
  • 8. THE SAME CONCEPT CAN BE APPLIED TO OTHER SCENARIOS
  • 9. INPUTs : Images of objects of interests Feed to NN Object shapes BUILD A MODEL TO LEARN VIDEO SEMANTIC SEGMENTATION (ENET) Model Prediction: The Neural Networks
  • 10. INPUTs : Images of objects of interests Feed to NN Object labels + location BUILD A MODEL TO LEARN OBJECT DETECTION IN VIDEO (SSD) Model Prediction: The Neural Networks
  • 11. Raw input abdominal CT scan Output segment BUILD A MODEL TO SEGMENT 4D CT SCAN ORGANS (NIFTYNET)
  • 12. BUILD A MODEL TO LEARN TO PLAY ANGRY BIRD

Editor's Notes

  • #5: https://www.kaggle.com/kmader/electron-microscopy-3d-segmentation/home
  • #6: https://www.kaggle.com/kmader/electron-microscopy-3d-segmentation/home
  • #11: https://www.kaggle.com/kmader/electron-microscopy-3d-segmentation/home
  • #12: https://www.kaggle.com/kmader/electron-microscopy-3d-segmentation/home
  • #13: https://www.kaggle.com/kmader/electron-microscopy-3d-segmentation/home