Slide 1

Slide 1 text

June 18, 2020

Slide 2

Slide 2 text

GDG Seattle GDG Seattle is an inclusive community where developers, designers, and entrepreneurs of all skill levels, genders, religions, and backgrounds are welcome to learn, practice, and share Google technologies, services, and platforms. Our motto is “be excellent to each other”. If you see or experience anything different, please contact GDG Seattle organizers.

Slide 3

Slide 3 text

@AndrewFerlitsch | @margaretmz | @xCatG | @GDGSeattle | #TensorFlow 2.x | #Keras | #DeepLearning | #ComputerVision Agenda ● Instructors intro ● Session 4 Review ● Q&A ● Next steps for session 5 Post questions to YouTube live stream & doc bit.ly/dldp-qa Tweet with hashtags below. 3 Agenda

Slide 4

Slide 4 text

@AndrewFerlitsch | @margaretmz | @xCatG | @GDGSeattle | #TensorFlow 2.x | #Keras | #DeepLearning | #ComputerVision Instructor - Andrew Ferlitsch ● Google Cloud AI / Developer Relations ○ Conferences ○ Universities ○ Meetups ○ Training / Workshops ○ Enterprise Clients ● Expertise is Computer Vision and Deep Learning ○ Formerly principal research (imaging) scientist for 20 yrs. ○ 115 US issued patents based on my research ● Writing a Deep Learning book to be published by Manning 4

Slide 5

Slide 5 text

@AndrewFerlitsch | @margaretmz | @xCatG | @GDGSeattle | #TensorFlow 2.x | #Keras | #DeepLearning | #ComputerVision Instructor / organizer - Margaret ● ML GDE (Google Developer Expert) ● Leading GDG Seattle, and Seattle Data/Analytics/ML ● Working on AI for art & design 5

Slide 6

Slide 6 text

@AndrewFerlitsch | @margaretmz | @xCatG | @GDGSeattle | #TensorFlow 2.x | #Keras | #DeepLearning | #ComputerVision GDG Seattle organizer - Yenchi ● Android Software Developer for 10+ years ● Co-organizer of GDG Seattle ● Did a few projects with ML pipeline 6

Slide 7

Slide 7 text

@AndrewFerlitsch | @margaretmz | @xCatG | @GDGSeattle | #TensorFlow 2.x | #Keras | #DeepLearning | #ComputerVision Please study prior to session 4 (6/18/2020): 7 Topics Video Slides Workshop AutoEncoders 45 min talk + 15 min workshop Link Link Link (to be added) Hyperparameter Tuning 45 min talk + 15 min workshop Link Link Link (to be added) Transfer Learning 45 min talk + 15 min workshop Link Link Link (to be added) You should have studied

Slide 8

Slide 8 text

@AndrewFerlitsch | @margaretmz | @xCatG | @GDGSeattle | #TensorFlow 2.x | #Keras | #DeepLearning | #ComputerVision 8 Review - Autoencoder Learned Pooling Learned Pooling Learned Pooling ... Learned Unpooling Learned Unpooling .. . Encoding Learning Latent Space x’ Stem Learned Unpooling Learner Reconstruction Decoding Learning Representational Learning Transformation Learning

Slide 9

Slide 9 text

@AndrewFerlitsch | @margaretmz | @xCatG | @GDGSeattle | #TensorFlow 2.x | #Keras | #DeepLearning | #ComputerVision 9 Review: Hyperparameter Tuning Learning Rate Scheduler Model Training Pre-Training Model Instances Warmup Model Training Model Instances Model Training Model Training Model Instances Warmup weight initialization using very small learning rate on ensemble of models. Select model instance with best numerical stability (lottery principle). Model instance with the selected weight initialization. Learn the optimal learning rate schedule Model instance with the selected weight initialization. Full training with the learned learning rate schedule.

Slide 10

Slide 10 text

@AndrewFerlitsch | @margaretmz | @xCatG | @GDGSeattle | #TensorFlow 2.x | #Keras | #DeepLearning | #ComputerVision 10 Review: Transfer Learning Stem Convolution Group Conv/ Residual Group Final Pooling Flatten Layer Conv/ Residual Group Conv/ Residual Group Classifi er (Dense) Layer 1 2 3 4 5 Coarse Level Training Fine-Tuning Training

Slide 11

Slide 11 text

@AndrewFerlitsch | @margaretmz | @xCatG | @GDGSeattle | #TensorFlow 2.x | #Keras | #DeepLearning | #ComputerVision 11 Margaret’s Transfer Learning Course

Slide 12

Slide 12 text

@AndrewFerlitsch | @margaretmz | @xCatG | @GDGSeattle | #TensorFlow 2.x | #Keras | #DeepLearning | #ComputerVision 12 Open Q&A Ask Your Questions

Slide 13

Slide 13 text

@AndrewFerlitsch | @margaretmz | @xCatG | @GDGSeattle | #TensorFlow 2.x | #Keras | #DeepLearning | #ComputerVision How do you study? ● Read Andy’s DL Primer (link) ● Watch the videos (bit.ly/2z3kbIA ) ● Do the exercises (bit.ly/2zMRVdw ) ● Join slack to help each other out (bit.ly/ml-study-jams) ● Bring your questions to Study Jam sessions for live Q&A 13

Slide 14

Slide 14 text

@AndrewFerlitsch | @margaretmz | @xCatG | @GDGSeattle | #TensorFlow 2.x | #Keras | #DeepLearning | #ComputerVision Please study prior to session 5 (7/9/2020): 14 Topics Video Slides Data Distributions 1hr talk + 30 min workshop Link Link Data Pipeline #1 45 min talk Link Link Data Pipeline #2 45 min talk Link Link What’s next?

Slide 15

Slide 15 text

@AndrewFerlitsch | @margaretmz | @xCatG | @GDGSeattle | #TensorFlow 2.x | #Keras | #DeepLearning | #ComputerVision Missed this session? Catch up with YouTube recording here. Happy studying & see you on 7/9! 15 Thanks for joining us!