Transfer Learning
for IoT
GEETA CHAUHAN JUNE, 2017
Intelligent Door
Welcome Sara
Good Kitty
What is
Deep
Learning?
 AI Neural Networks
composed of many
layers
 Learn like humans
 Automated Feature
Learning
 Layers are like Image
Filters
Tensorflow
Android Things
 Add Tensorflow Inferencing
library
 Create Inference object,
load model
 feed() image, run() to predict
 fetch() to get classification
result
Transfer
Learning
Transfer Learning – New Classifier
Inception V3
Optimizations
 Graph Transform Tool
 Freeze graph (variables to constants)
 Quantize weights (20 M weights for IV3)
 Quantization (32 bit float → 8 bit float)
 Memory Mapping
 Inception v3 93 MB → 1.5 MB
Intelligent Door Bell
Intel Edison + Seed Studio Kit Raspberry Pi + Pi Camera
Tensorflow IoT Pipeline
Inference on DeviceTrain Model in Cloud
Cucumber
Sorter
Caltrain Rider
 Realtime Caltrain arrival
prediction
 Audio Visual pipeline on
Raspberry pi
 Image classification for
Caltrains
And Some More …
Traffic sign
detector for
Self Driving Car
Wearable
assistant for
Blind (Horus)
Smart cameras,
smart door lock
Human line
counter
Real time
exercise score
on smart watch
Creative Arts,
Music
Generators
Intelligent
Robots
Look into the Future
 Hardware: Neural Network Chips
Intel Fathom Neural Stick Nvidia Jetson
References
 Tensorflow for Android Things Sample
 Tensorboard hands-on
 Graph Transform Tool
 Tensorflow for Poets (Transfer learning)
 Tensorflow for Poets2 (Optimizations for Mobile/IoT)
 Cucumber Farmer Deep Learning Story
 Caltrain Rider story
 Intel Fathom Neural Stick
 Nvidia Jetson
Questions?
Contact
https://www.linkedin.com/
in/geetachauhan/
geeta@svsg.co
Common Problem Solutions
 Tensorboard is your friend – X-Ray vision
 Image size mismatch for the input tensor
 Output classifier specify correct number of classes from your model
 Model too large – Load in memory
 Missing ops -- Graph transform tool
 Device heating up under heavy processing load
 Split the model, do part detection on device, rest in cloud
 Reduce the frequency eg only do on movement detection
Era of AI First
 Billions of connected devices
 Intelligence at the Edge
 Increasing Computation power
 Edison: 500 MHz, 1 GB RAM
 RPi3: 1.2 GHz Quad-core
 Deep neural networks running
on the IoT device
 Local inferencing →
compressed insights to cloud

Transfer learning for IoT

  • 1.
  • 2.
  • 3.
    What is Deep Learning?  AINeural Networks composed of many layers  Learn like humans  Automated Feature Learning  Layers are like Image Filters
  • 4.
    Tensorflow Android Things  AddTensorflow Inferencing library  Create Inference object, load model  feed() image, run() to predict  fetch() to get classification result
  • 5.
    Transfer Learning Transfer Learning –New Classifier Inception V3
  • 6.
    Optimizations  Graph TransformTool  Freeze graph (variables to constants)  Quantize weights (20 M weights for IV3)  Quantization (32 bit float → 8 bit float)  Memory Mapping  Inception v3 93 MB → 1.5 MB
  • 7.
    Intelligent Door Bell IntelEdison + Seed Studio Kit Raspberry Pi + Pi Camera
  • 8.
    Tensorflow IoT Pipeline Inferenceon DeviceTrain Model in Cloud
  • 9.
  • 10.
    Caltrain Rider  RealtimeCaltrain arrival prediction  Audio Visual pipeline on Raspberry pi  Image classification for Caltrains
  • 11.
    And Some More… Traffic sign detector for Self Driving Car Wearable assistant for Blind (Horus) Smart cameras, smart door lock Human line counter Real time exercise score on smart watch Creative Arts, Music Generators Intelligent Robots
  • 12.
    Look into theFuture  Hardware: Neural Network Chips Intel Fathom Neural Stick Nvidia Jetson
  • 13.
    References  Tensorflow forAndroid Things Sample  Tensorboard hands-on  Graph Transform Tool  Tensorflow for Poets (Transfer learning)  Tensorflow for Poets2 (Optimizations for Mobile/IoT)  Cucumber Farmer Deep Learning Story  Caltrain Rider story  Intel Fathom Neural Stick  Nvidia Jetson
  • 14.
  • 15.
    Common Problem Solutions Tensorboard is your friend – X-Ray vision  Image size mismatch for the input tensor  Output classifier specify correct number of classes from your model  Model too large – Load in memory  Missing ops -- Graph transform tool  Device heating up under heavy processing load  Split the model, do part detection on device, rest in cloud  Reduce the frequency eg only do on movement detection
  • 16.
    Era of AIFirst  Billions of connected devices  Intelligence at the Edge  Increasing Computation power  Edison: 500 MHz, 1 GB RAM  RPi3: 1.2 GHz Quad-core  Deep neural networks running on the IoT device  Local inferencing → compressed insights to cloud