Big Data Days 2019

 8-10 октября   Москва

Constant Bridon

OCTO Technology, France

Синхронный перевод

Биография

Constant has been a consultant of Octo Technology’s Data Science tribe for more than three years. He graduated from Ecole Normale Supérieure, where he developed a strong experience in applied research : following his graduation, he co-signed 4 articles published in specialised conferences. Two are related to Telecommunications, whereas the other 2 deal with Human-Machines interfaces.
With a master’s degree major in applied mathematics, Constant spent a year at NICTA (now Data61 part of Csiro) in Sydney, working on Data Science related topics within the Machine Learning Research Group.
Because he is concerned about sharing his knowledge, Constant interweaves his consulting activity with recurrent training missions thanks to Octo’s training organisation, OCAC. He is responsible for several topics : Fundamentals of Data Science, Advanced Data Science, Pig & Hive, Spark with Python and Scala. He aims to become trainer AWS Big Data.
Besides, Constant leads a significant technological watch on two data science subjects : interpretation of ensemble models (gradient boosting), and deploying machine learning algorithm on dedicated hardware. 
Last, but not least, Constant has been taking part of Iron Car races since January 2018. The goal is to use a rasberry and a camera to make a toy car autonomous, using deep learning models.

Доклад

On Board Artificial Intelligence : Train, Deploy and Use Deep Learning on an Edge Device, a Raspberry Pi

Machine learning applications are new in the software development landscape, and tend to be hard to build. As Google noted in an article (source : https://papers.nips.cc/paper/5656-hidden-technical-debt-in-machine-learning-systems.pdf), it is mainly because the application is much broader than the model itself. Surprisingly though, Machine Learning applications follow a double Pareto’s law. On the one hand, 80% of the time spent on building those applications deals with machine learning problems whereas 20% of the remaining time is spent on integrating the model to a running application. On the other hand, only 20% of the code lines are specific to machine learning ; the vast rest is about integration and run.
I would like to first explore the foundations of this trend, to then show why it kills machine learning application development and sustainability.
In order to illustrate the tips and tricks of shipping a deep learning model to production, I would use a live demo of a model designed to recognised car drawings via the camera of a Raspberry Pi. It would allow me to. Furthermore, by identifying the main stages of the application life cycle (training, deploying and using), I will lay the emphasis on the common mistakes one does not bear in mind to make a successful machine learning product.

Ключевые слова

Deep Learning
Deployment
On Board
Production