Software 2.0 - Disruption

Presentation for KI University introducing the concept of Software 2.0, how it relates to the mass spread of ML, and how Tesla is disrupting the automotive industry.

Transcript

1. SOFTWARE 2.0 SIMPLE CARS WITH COMPLEX SOFTWARE JOSH FRIEDMAN 2018/09/20
2. OVERVIEW ROUGH CONTENTS ▸ What is Software 2.0? ▸ Deep Learning for the Masses ▸ Disruption in the Automotive Industry ▸ Future Implications ▸ What to Remember…
3. WHAT IS SOFTWARE 2.0?
4. WHAT IS SOFTWARE 2.0? SOFTWARE 1.0 ▸ Programmer decides desired behaviour ▸ Explicit instructions to the computer by the programmer ▸ Written in known languages - Python, Java, C++
5. WHAT IS SOFTWARE 2.0? SOFTWARE 2.0 ▸ Abstract ▸ Human unfriendly language ▸ Programmers instead specify the behaviour with a rough skeleton of the search space and optimise
6. 1.0 Programmers ▸ Maintain surroundings ▸ Tooling ▸ Analytics ▸ Visualizations ▸ Infrastructure 2.0 Programmers ▸ Maintain data for training ▸ Maintains ▸ Massage ▸ Clean ▸ Label WHAT IS SOFTWARE 2.0? SOFTWARE 2.0
7. ▸ What is the stack for Software 2.0? ▸ ____________________ ▸ What will be the VCS for Software 2.0? ▸ ____________________ WHAT IS SOFTWARE 2.0? SOFTWARE 2.0
8. DEEP LEARNING FOR THE MASSES
9. DEEP LEARNING FOR THE MASSES DEEP LEARNING AS A COMMODITY ▸ Training a deep neural network requires little to no understanding of numerical stability, cross entropy, etc… ▸ Open source and have enabled anyone to train and/or deploy a model ▸ … ▸ Which simplifies the picture to how to build the most effective training dataset?
10. DEEP LEARNING FOR THE MASSES DON’T DO TRY THIS AT HOME ▸ Connected RPi3 with camera pointed at the street ▸ Minimal effort to build a scheduled inference workflow ▸ Goal is to identify free parking spots*
11. DEEP LEARNING FOR THE MASSES DON’T DO TRY THIS AT HOME ▸ Connected RPi3 with camera pointed at the street ▸ Minimal effort to build a scheduled inference workflow ▸ Goal is to identify free parking spots*
12. DATA IS THE NEW OIL. IT’S VALUABLE, BUT IF UNREFINED IT CANNOT REALLY BE USED. IT HAS TO BE CHANGED INTO GAS, PLASTIC, CHEMICALS, ETC TO CREATE A VALUABLE ENTITY THAT DRIVES PROFITABLE ACTIVITY. Clive Humby (2006) DEEP LEARNING FOR THE MASSES
13. DEEP LEARNING FOR THE MASSES LABELS ARE INVALUABLE
14. TESLA AUTOMATIC WINDSHIELD WIPERS ▸ Dashboard cameras from the entire Tesla fleet results in a massive collection of data ▸ Essentially driving data streams ▸ Labelling is complex with real-world conditions DEEP LEARNING FOR THE MASSES
15. REAL WORLD TESTING DEEP LEARNING FOR THE MASSES
16. DISRUPTION IN THE AUTOMOTIVE INDUSTRY
17. DISRUPTION DISRUPTION IN THE AUTOMOTIVE INDUSTRY
18. STATUS QUO DISRUPTION IN THE AUTOMOTIVE INDUSTRY ▸ Standard car companies fight for “dashboard” space ▸ Steering wheel team hates the entertainment team ▸ Electric and electronics are separate and independent ▸ This results in a complex car with simple firmware ▸ Tesla flipped this “equation” on its head ▸ Simple cars with complex software
19. WHAT IS TESLA DOING? DISRUPTION IN THE AUTOMOTIVE INDUSTRY
20. REAL DISRUPTION DISRUPTION IN THE AUTOMOTIVE INDUSTRY ▸ The first reaction - “Let’s just hire some developers!” ▸ [internal mess] ▸ The next solution - “Let’s just buy it”
21. SOFTWARE 2.0 IN ACTION DISRUPTION IN THE AUTOMOTIVE INDUSTRY ▸ Tesla is not competing with Detroit ▸ Tesla is now battling with Silicon Valley for automation ▸ Its vast fleet equipped with sensors is their New New Oil ▸ Sell cars with sensors ▸ Collect more sensor data ▸ Train better model ▸ Push OTA autonomy upgrades ▸ Sell more cars with sensors
22. FUTURE IMPLICATIONS
23. FUTURE IMPLICATIONS THE PATH TO AUTONOMY Tesla’s Conceptual Software “Divide” over Time
24. FUTURE IMPLICATIONS PRIVACY ▸ Autonomous cars powered by Software 2.0 will enable mass amounts of continuous, high definition 360° video
25. FUTURE IMPLICATIONS PRIVACY Prepare yourself for even more targeted advertisements, and good luck “opting” out!
26. FUTURE IMPLICATIONS MANDATORY BLACK MIRROR REFERENCE
27. WHAT TO REMEMBER
28. CONCLUSION A FEW POINTS ▸ Software divide exists for ML applications ▸ Maintain surroundings vs. maintain data for training ▸ Deep Learning is now a commodity ▸ Training data >>>>>> understanding model architecture ▸ Tesla ships simple cars with complex software ▸ Automotive training data is invaluable for success ▸ Don’t forget about future privacy issues!
30. REFERENCES LIST OF SOURCES ▸ Building the Software 2.0 Stack ▸ Labeled Data ▸ Software Disruption in the Automotive Industry ▸ Tesla’s Model 3 UI ▸ Move Fast and Break Things ▸ Autonomous Vehicles and Privacy