Democratization of AI: from Researchers to Any User


Beaucoup de sociétés de réfléchissent à leur propre évolution dans un monde de plus en plus digital et le machine learning est devenu incontournable. Ces techniques techniques existent depuis longtemps et elles se sont récemment démocratisées. Cette présentation revient sur les ingrédients qui ont permis leur popularisation.

Machine learning somehow became one magic wand able to solve many things. It can be seen as a way to automate an existing pipeline which became very time consuming, it could also be seen as a way to process information we could not handle so far such as images or videos. We usually forget it is first a statistical method which is right in most cases but not in all cases. This talk will expose a particular case where machine learning and deep learning will be used to reduce the amount of manual work.

Democratization of AI: from Researchers to Any User

Key Ideas

Many things which did exists ten years ago were not as easy to use as they are now. It was more like a researcher’s job than an engineer’s job. That’s something which changes in the past five years because four pieces became available at the same time.

Democratization of algorithms

Open Source became more and more popular, github without being very new was a tremondous catalyst as it hosts the vast majority of open source projects. Anybody can find an implementation of algorithms described in research papers also available through initiative such as arxiv or jmlr. Most of these algorithms are now available in Python.

Democratization of hardware

The technology improved but also the economical model. Computing power can still be acquired by getting machines or cluster of machines but it can also be rent for a short period of time.

Democratization of AI

By AI, I mostly mean deep learning. Difficult to optimize and very expensive to train ten years ago, it is now more accessible. Frameworks are open source such as :epkg:`CNTK` or :epkg:`pytorch` and many trained models are available. Competing into a challenge such as ImageNet require a lot computing power not everybody can afford. The winner of Image Net 2015 was Microsoft Research Asia. But these trained models can be leveraged with transfer learning and adapted to a new similar problem without paying the cost of a full training process.

Democratization of knowledge

The first reflex is now to go to internet to find an answer to a technical question. Stackoverflow or Quora have gathered so many answers that they pop up in first position to many queries sent to a search engine. People share technical knowledge. Teachers also share theoritical knowledge through videos, tutorials on Coursera, EdX or Fun MOOC (France). Even the documentation of libraries is not only a list of functions but also contains many examples and tutorials users can copy paste and adapt to their own problems.