[LUM#22] I came, I saw, industry
It is the year 2024 A.D. Every sector has been taken over by artificial intelligence… Every sector? No! The indomitable manufacturing industry continues to resist this revolution. A resistance that mathematician Bijan Mohammadi might just overcome thanks to his magic formula.
He is neither a druid nor a warrior, but a researcher at the Alexander Grothendieck Institute in Montpellier. By swapping his cauldron for a computer and his potions for an algorithm, Bijan Mohammadi has developed an AI system that is unique in the world and convinced manufacturers such as Renault of its value. “It uses physical signals such as noise or vibrations to detect anomalies and perform preventive maintenance, ” he explains.
Hearing the snap of a tiny rubber connector to identify a defective car, or detecting wear on a milling cutter head by monitoring changes in vibration—these are the kinds of tasks this AI is capable of. “No big deal, ” some might say, except that… “This allows us to produce up to 300% more parts with the same milling head, which represents a savings of several hundred thousand euros per year per production line,” emphasizes the mathematician , “who is no stranger to the industrial world and its logic.”
AI and Industry: The Rift
Savings in money, time, and quality. That was all it took to convince this sector, because while one might imagine AI reigning over the industry like Caesar over Gaul, that is far from the case. There are many reasons for this reluctance, starting with the lack of robustness in traditional AI systems, whose rigid architectures are not designed to adapt to the variability of industrial sound environments. “We see this with voice assistants on phones that glitch if the user has a strong accent or if there’s background noise,” explains Bijan Mohammadi.
Added to this is the issue of data. For reasons of sovereignty, manufacturers refuse to turn to GAFAM to share the data needed to train AI. There remains the question of scaling up to an industrial level and the lack of proof of return on investment—in short… “Many difficulties, few guarantees—an equation that doesn’t sit well with the business world,” sums up the mathematician, who has therefore taken a gamble on a different approach.
From Caesar's Laurels to Bondzai
Summer 2016 AD, while Gaul was turning in on itself, Bijan Mohammadi was hard at work on his 12 labors and developing the algorithm that would open the realm of the gods to him. This algorithm replaces the fixed architecture of classical AIs—which consists of a fixed and generally enormous number of neural layers—with a neural generator that stacks layers in real time, much like a 3D printer. “It’s a kind of miniature AI that creates its own architecture. The advantage is twofold: it requires very few examples and adapts completely to the environment in which it is placed. It is very reliable since it involves only a few specific parameters.”
And since there’s little data, there’s no cloud—so no privacy concerns and no massive energy costs either. “It’s a frugal AI that runs on standard processors requiring very little energy and no investment in IT infrastructure.” Reliability , adaptability, cost-effectiveness… These are the ingredients of the success marketed since 2021 by the startup Bondzai, for which Bijan Mohammadi(x) continues to reap the rewards.
UM podcasts are now available on your favorite platform (Spotify, Deezer, Apple Podcasts, Amazon Music, etc.).