Part of Microblink’s AI team flew to southern France to attend the first edition of the World Artificial Intelligence Cannes Festival (WAICF).
The event, held from April 14th to 16th, took place in the renowned Palais des Festivals, where more than a hundred companies showcased their AI expertise through workshops, presentations and start-up pitches. Here are some of the main highlights from the event.
Using machine learning to predict traffic accidents
Michelin apparently does more than rate restaurants and make tires. The French mobility giant launched an internal startup called DDI (Driving Data Intelligence) to analyze driving behavior and identify high risk hotspots on the road, such as dangerous curves.
Philippe Geneste, CTO at DDI, talked about how his team uses geographical, weather and driving data — a lot of which is sourced from Michelin’s fleet management services — to predict hazards. DDI engineers go so far as to look at seasonal vegetation that may impair a driver’s visibility when taking a turn, using computer vision to analyze satellite imagery and monitor vegetation density. These insights are then shared with road authorities who are able to plan their road maintenance more efficiently.
Processing language without an intermediary
NLP and conversational AI were a recurring theme at the conference, both in the talks and exhibitor booths.
A talk by Meta, focused on direct text translation, was particularly interesting. Antoine Bordes, Meta AI’s Managing Director, showed how current methods of translating through English cause information to be lost in the process. Meta has built models for true, language-to-language translation without English in between. They’re also working on speech-to-speech translation without transcribing text and converting it back to speech. While this is still in the works, it has the potential to retain crucial elements of speech, including tone, emphasis, emotion, rhythm and more.
Rethinking the quality of AI
On the start-up stage, Alex Combessie introduced attendees to the importance of Quality Assurance (QA) in AI.
Giskard’s co-founder and CEO did a quick demo of the product, correcting the output of a model predicting whether a user would default on the loan they’re applying to take out. It all starts in your usual prototyping environment — the Python notebook. This is where Giskard will generate a visual inspector to help you and your team members test the way your models behave on individual cases. In the below example, a human is telling the model that the fact the applicant is a woman shouldn’t affect the prediction. This can prevent algorithmic bias from occurring in production and make the model more robust in its decision making.
Finding leaked passwords on Github.
A demo hosted by Slim Trabelsi from Credential Digger presented an ML-based approach to scanning Github repositories for exposed credentials such as passwords, API keys, signatures, hashed data and IP addresses.
Unlike other scanning tools (including Github’s official scanner), Credential Digger doesn’t suffer from a high false positive rate thanks to two machine learning models it’s running under the hood:
- File Path Model
This model is in charge of analyzing the path and file naming of a project to filter out all the dummy credentials that can be identified as hits by traditional scanners.
- Code Snippet Model
This model identifies portions of code that perform a real authentication action and is able to distinguish between real and fake credentials at a reduced false positive rate.
AI as a force for good
Overall, the WAICF has been an exciting experience with plenty of interesting companies and talks. The conference is oriented more towards business than tech, but it still provides valuable insights into practical uses of AI in a way that promotes positive social impact.
Interested to see how we approach AI at Microblink? Learn more.