Expert systems, transfer learning and its impact on
your next big project
Galvin Widjaja
Lauretta.io
Galvin Widjaja
@galvinw | github.com/galvinw
Graduated: Singapore Management University is Quant Finance
Work:
Management Consulting → Business Process Management → Monte Carlo
Systems → Process Strategy → IT Strategy → Data Science
Transition:
MIT Visiting Scholar → CFO at Pinchfavor Inc. (New York)
Current:
Founder - Lauretta.io (ML Pipeline and Human Automation)
CEO - Bttrigitical AI (Petrochem AI Solutions)
Management Director - Es Teler 77 Singapore
Indonesia Developer Summit 2017
About Me
Agenda
● What is transfer learning, what are expert systems and are robots going to take over the world
● Build your own image recognition in 20 minutes
● The Limitations for Small timers and why the limits are being lifted
Indonesia Developer Summit 2017
Indonesia Developer Summit 2017
Transfer Learning
A field focused on storing knowledge
gained while solving one problem and
applying it to a different but related
problem
Indonesia Developer Summit 2017
Indonesia Developer Summit 2017
Indonesia Developer Summit 2017
Expert Systems
A computer system that emulates the
decision-making ability of a human
expert
Indonesia Developer Summit 2017
Expert Systems
Are machines going to take over the world?
Well, obviously they are...
Indonesia Developer Summit 2017
So you better be on their good side when it happens
Build your own image recognition in 20 minutes
Indonesia Developer Summit 2017
Indonesia Developer Summit 2017All code at : https://github.com/galvinw/deeplearning.ai thanks to Deeplearning.ai
Initialize
Python with Keras, Tensorflow and Numpy. Essentially this allows you to do
any image recognition task
Indonesia Developer Summit 2017All code at : https://github.com/galvinw/deeplearning.ai thanks to Deeplearning.ai
Define a model: Extract Features
Where are the:
1. Vertical lines
2. Horizonal lines
3. Circles
4. Big circles
5. Clear.. circles?
6. ...squiggles
Indonesia Developer Summit 2017All code at : https://github.com/galvinw/deeplearning.ai thanks to Deeplearning.ai
Define a model: Analyse with DNN
Each feature becomes a set of arbitrary data points that is stored in a large set of
neurons
Indonesia Developer Summit 2017All code at : https://github.com/galvinw/deeplearning.ai thanks to Deeplearning.ai
Looks super hard… but only 9 lines of code
Thanks to Keras (High level framework) and Tensorflow (Handles all the
optimization and complicatedness)
This line says, use
32 types of
features.
Optional: Recognise the object first
Indonesia Developer Summit 2017
This is hard (we won’t touch it)
Indonesia Developer Summit 2017All code at : https://github.com/galvinw/deeplearning.ai thanks to Deeplearning.ai
Initialize
Python with Keras, Tensorflow and Numpy. Essentially this allows you to do
any image recognition task
Indonesia Developer Summit 2017All code at : https://github.com/galvinw/deeplearning.ai thanks to Deeplearning.ai
Identify Objects
Basically you draw boxes all over an
image and check if there’s something
interesting in one of them
Indonesia Developer Summit 2017All code at : https://github.com/galvinw/deeplearning.ai thanks to Deeplearning.ai
Suppress bad predictions
Called “intersection over union
(IoU)” This merges good bounding
boxes
Indonesia Developer Summit 2017All code at : https://github.com/galvinw/deeplearning.ai thanks to Deeplearning.ai
Suppress bad predictions
First, you choose boxes that don’t
have high likelihood of having a
feature and discard it, then you
merge the rest of the boxes using
“intersection over union (IoU)”.
- This merges good bounding boxes
Indonesia Developer Summit 2017All code at : https://github.com/galvinw/deeplearning.ai thanks to Deeplearning.ai
Do Image Recognition of each box
First, you choose boxes that don’t have high likelihood of having a feature
and discard it, then you merge the rest of the boxes using “intersection
over union (IoU)”.
- This merges good bounding boxes
The Limitations for Small timers
and why the limits are being lifted
Indonesia Developer Summit 2017
Indonesia Developer Summit 2017
The Venn diagram of our problems
No Data
Complicated
Indonesia Developer Summit 2017
Not with these
frameworks
Tensorboard and
development is available
instantly with the use of
tensorflow
Lacking Data?
Indonesia Developer Summit 2017
Transfer learning is fully proven to work well on all image recognition.
Most frameworks are trained on ImageNet given you a baseline of 1 million images
Also new techniques seem to be able to use up to 1% as much data as previously required
Processing Power
Indonesia Developer Summit 2017
With smaller data requirements, processing needs
are decreasing. But even so, Google ML provides
fairly cheap GPU usage with many student
concessions
Conclusion
The robots are coming to be our overlords.
Get a job as one of their technicians now.
(No seriously, get ready for this)
… also we might be recruiting
Indonesia Developer Summit 2017

Slides galvin-widjaja

  • 1.
    Expert systems, transferlearning and its impact on your next big project Galvin Widjaja Lauretta.io
  • 2.
    Galvin Widjaja @galvinw |github.com/galvinw Graduated: Singapore Management University is Quant Finance Work: Management Consulting → Business Process Management → Monte Carlo Systems → Process Strategy → IT Strategy → Data Science Transition: MIT Visiting Scholar → CFO at Pinchfavor Inc. (New York) Current: Founder - Lauretta.io (ML Pipeline and Human Automation) CEO - Bttrigitical AI (Petrochem AI Solutions) Management Director - Es Teler 77 Singapore Indonesia Developer Summit 2017 About Me
  • 3.
    Agenda ● What istransfer learning, what are expert systems and are robots going to take over the world ● Build your own image recognition in 20 minutes ● The Limitations for Small timers and why the limits are being lifted Indonesia Developer Summit 2017
  • 4.
    Indonesia Developer Summit2017 Transfer Learning A field focused on storing knowledge gained while solving one problem and applying it to a different but related problem
  • 5.
  • 6.
  • 7.
    Indonesia Developer Summit2017 Expert Systems A computer system that emulates the decision-making ability of a human expert
  • 8.
    Indonesia Developer Summit2017 Expert Systems
  • 9.
    Are machines goingto take over the world? Well, obviously they are... Indonesia Developer Summit 2017 So you better be on their good side when it happens
  • 10.
    Build your ownimage recognition in 20 minutes Indonesia Developer Summit 2017
  • 11.
    Indonesia Developer Summit2017All code at : https://github.com/galvinw/deeplearning.ai thanks to Deeplearning.ai Initialize Python with Keras, Tensorflow and Numpy. Essentially this allows you to do any image recognition task
  • 12.
    Indonesia Developer Summit2017All code at : https://github.com/galvinw/deeplearning.ai thanks to Deeplearning.ai Define a model: Extract Features Where are the: 1. Vertical lines 2. Horizonal lines 3. Circles 4. Big circles 5. Clear.. circles? 6. ...squiggles
  • 13.
    Indonesia Developer Summit2017All code at : https://github.com/galvinw/deeplearning.ai thanks to Deeplearning.ai Define a model: Analyse with DNN Each feature becomes a set of arbitrary data points that is stored in a large set of neurons
  • 14.
    Indonesia Developer Summit2017All code at : https://github.com/galvinw/deeplearning.ai thanks to Deeplearning.ai Looks super hard… but only 9 lines of code Thanks to Keras (High level framework) and Tensorflow (Handles all the optimization and complicatedness) This line says, use 32 types of features.
  • 15.
    Optional: Recognise theobject first Indonesia Developer Summit 2017 This is hard (we won’t touch it)
  • 16.
    Indonesia Developer Summit2017All code at : https://github.com/galvinw/deeplearning.ai thanks to Deeplearning.ai Initialize Python with Keras, Tensorflow and Numpy. Essentially this allows you to do any image recognition task
  • 17.
    Indonesia Developer Summit2017All code at : https://github.com/galvinw/deeplearning.ai thanks to Deeplearning.ai Identify Objects Basically you draw boxes all over an image and check if there’s something interesting in one of them
  • 18.
    Indonesia Developer Summit2017All code at : https://github.com/galvinw/deeplearning.ai thanks to Deeplearning.ai Suppress bad predictions Called “intersection over union (IoU)” This merges good bounding boxes
  • 19.
    Indonesia Developer Summit2017All code at : https://github.com/galvinw/deeplearning.ai thanks to Deeplearning.ai Suppress bad predictions First, you choose boxes that don’t have high likelihood of having a feature and discard it, then you merge the rest of the boxes using “intersection over union (IoU)”. - This merges good bounding boxes
  • 20.
    Indonesia Developer Summit2017All code at : https://github.com/galvinw/deeplearning.ai thanks to Deeplearning.ai Do Image Recognition of each box First, you choose boxes that don’t have high likelihood of having a feature and discard it, then you merge the rest of the boxes using “intersection over union (IoU)”. - This merges good bounding boxes
  • 21.
    The Limitations forSmall timers and why the limits are being lifted Indonesia Developer Summit 2017
  • 22.
    Indonesia Developer Summit2017 The Venn diagram of our problems No Data
  • 23.
    Complicated Indonesia Developer Summit2017 Not with these frameworks Tensorboard and development is available instantly with the use of tensorflow
  • 24.
    Lacking Data? Indonesia DeveloperSummit 2017 Transfer learning is fully proven to work well on all image recognition. Most frameworks are trained on ImageNet given you a baseline of 1 million images Also new techniques seem to be able to use up to 1% as much data as previously required
  • 25.
    Processing Power Indonesia DeveloperSummit 2017 With smaller data requirements, processing needs are decreasing. But even so, Google ML provides fairly cheap GPU usage with many student concessions
  • 26.
    Conclusion The robots arecoming to be our overlords. Get a job as one of their technicians now. (No seriously, get ready for this) … also we might be recruiting Indonesia Developer Summit 2017