Get these slides!

​

http://bit.ly/craft-ai-APIDaysParis

and

@craft_ai

Trust, Accountability and DX: cracking AI challenges with White-box Machine Learning

​

​

Paris API Days - 14/12/16

Clodéric Mars - CTO @ craft ai

9 experts in Artificial Intelligence

Spin-off from MASA Group & initial funding from Talis in June 2015

Beta released in April 2016

Public release & first projects in production S1 2017

The Promise

More data, more APIs should make our life easier

make us more productive

The Reality

More data / APIs means more complexity

needs programming / curation

at craft ai we

learn how a system is used, continuously,

to automate it

to make recommendation

to detect anomalies

craft ai is a hosted machine learning API that delivers actionable decision models from each user activity and context history, in real-time.

Benefits

Contextualized

White-box �Machine Learning

User-centric

...

Adaptable UX

Health & Wellness

Personalized Coach

​

Connected Things

Smart Automation

Utilities & Industry 4.0

Business Assistant

Conversational UI

Proactive Bot

White-box �Machine Learning

What is White-box ML?

Decisions are explainable�See also Explainable AI by Darpa�http://www.darpa.mil/program/explainable-artificial-intelligence

https://www.wired.com/2016/10/understanding-artificial-intelligence-decisions/

White-box ML

• Who? What for?
• Developers, to build and debug
• Regulators, to audit
• End users, to understand and trust
• Which kind of AI is concerned?
• How can we maintain a high level of accuracy?

How we do it at craft ai

Why us?

• Background
• Games & Simulation
• VFX & Animation
• Robotics
• Where AI needs to be “directed”
• by business experts
• by artists

At craft ai, we use Decision Trees!

they are debuggable

they give a reason for each decisions

they can easily trace decisions back to original data

Wait, don’t decision trees sucks?

they overfit

they produce low quality results

Well it’s not just about decisions trees

Thanks to our design

Feedback loop

• The “usual” process is
• Send actions & context change as they occur
• Compute decision tree
• Use decision tree to automate actions
• Users continuously provide feedback
• Reinforcement
• Contradiction

➡ Reduce unwanted overfitting

Thanks to our verticalization

Data types specialization

• Specialized split strategies
• Time data
• Geographical data
• ...

➡ Prediction improvement

Thanks to our R&D

Forgetting

• Over time, non-meaningful data are removed
• Users can change habits
• Enable incremental build

➡ Less overfitting�

D’amato A., Boussard M.: “Forgetting Methods for White Box Learning” PAAMS 2016

Takeaways on White-box ML

• Challenges
• Not a lot of research
• “Fancy” methods are not applicable
• Manage expectation
• Opportunities
• Automate knowledge work
• Increase trust of assistants
• Enforce regulations?

Artificial Intelligence in �Creative Industries

​

​

Past speakers: Watson, Pixar, Ubisoft, Google, MPC, INRIA, DeepMind, Blizzard, …

​

​

3rd edition - July 2017

Signup at beta.craft.ai

​

Follow us at @craft_ai

​

Join the team at craft.ai/jobs