The year 2018 is seeing the proliferation of AI business applications. Marketing, maintenance, logistics, human resources … all activities will be affected by the new technology. Our advice? Don’t miss this revolution.
The tale has been running in the French pharmaceutical industry for a few weeks now. “The head of a large French pharmaceutical laboratory was to reposition a molecule late in development, but the correct results were not obtained. He has pulled his team and created an artificial intelligence solution based on neural networks: Within 48 hours AI had given him a response. A response that would have taken a human three weeks to configure.
The story — which has not been verified — is indicative of not only the madness, but the fantasies that surround this amorphous concept of AI. What exactly can we expect from artificial intelligence? What services can it already offer to businesses? What are the offers available on the market?
All specialists are looking to the imminent takeoff of concrete applications of AI.
“The year 2018 is going to see the theory of AI seep into the practical world,” adds Rand Hindi, founder of PDG Snips, a Parisian start-up specializing in voice recognition technology. He claims to have been contacted by 300 contacts since the technology launched last June.
But what is AI today? Many industry leaders still have trouble pinpointing it exactly: “With AI, there are these systems that try to get as close as possible to human intelligence,” explains Romain Picard, Director for Southern Europe, the Middle East, and Cloudera Africa, a California-based firm specializing in Big Data software.
“AI is to delegate human capabilities to machines in terms of intelligence, decision making, and action,” says Mouloud Dey, director of Innovation at SAS France, (advanced analytics center.)
The difficulty in defining Artificial Intelligence stems simply from its history, during which several approaches have been attempted. But, evidently, a decisive step has been taking with machine learning, which is easier to implement. “With machine learning, we learn to operate the machine as a child,” Rand Hindi says. “Humans do not have to understand the phenomenon to teach the machine to reproduce it. In order to teach a neural network to recognize a cat image, for example, we show it images and tell it to recognize if the output is a cat or not. At no point should a man describe what a cat is.”
The End of Coding
“AI is the ‘end of code’, to quote Wired magazine’s famous title in 2016,” adds Jean-Phillipe Desbiolles, Vice President for France of Watson, the AI solution of IBM. “We are moving from a world of programming to a world of learning where knowledge, expertise, and life skills are key.”
Experts estimate that eventually, thanks to AI, we will be able to better detect and understand weak signals in medical imaging, satellite images, medical analysis, camera surveillance, tax declarations, customer complaints, social media. They also think that dialogue with computers will be done in a more and more natural language, (voice or writing with NLP- Natural Language Processing) or by intermediate images: The results will be represented with images, including 3D holograms.
Information is improving
“But today, concretely, AI is still very modest: we are not yet at disruptive applications, but on improving programming,” opines Julien Maldonato, consulting partner innovation at Deloitte. He and other colleagues at consulting firms believe that AI should be used primarily for document management (also called business process automation, or RPA – “Robotic Process Automation,” predictive management, logistics management, the sorting of resumes, the forecasting of sales and therefore of production, the responses to requests (purchases, requests for information, complaints, etc) that customers transmit by email or telephone, the detection of fraud.
The use cases are multiplying at the automotive supplier Faurecia, the establishment, with the help of Cloudera, of a network of sponsors connected to the factories has helped develop predictive maintenance and reduce halts in production.
Augustin Marty, co-founder and CEO of Deepomatic, a Parisian start-up specializing in image recognition, is working with Valeo’s Thermal System R&D Division, while also helping with the development of intelligent road tolling without barriers, (replaced by cameras) for one of the world infrastructure leaders. The French tax administration is also using AI on fraud detection.
Data Above All
What does it take to successfully implement an Artificial Intelligence application? Data, algorithms and humans. Data (from interactions with customers, internal processes, etc) is essential to the performance of artificial intelligence systems, thus highly strategic. “We need to ensure their quality and high accessibility within the company,” warns Vivien Tran-Thien, senior manager in the analytics team at EY.
Above all, the quantity and the nature of this data will ultimately guide the technological choices.”If you have a lot of data, you will be able to form your own application from the tools of major publishers, which are occasionally open source,” says Laurence Lafont, director of the Marketing & Operations division at Microsoft France. “If you have less data, you will have to work with a more specialized application.” Similarly, if the data is subject to constraints, such as personal data protection, they will be processed instead on the companies’ servers rather than the cloud.
The competition is already lively
And precisely, what’s needed on the technical side? Many variants are possible. Some publishers offer API’s (software bricks) available directly in their cloud for specific uses of artificial intelligence: “For example, there are Google Cloud APIs for image recognition, translation, speech-to-text (speech recognition to transcribe it in the form of machine readable text,)” recalls Gregoire Peiron, commercial director at Google Cloud in France.
In any case, the machine-learning tool remains at the heart of the system. Publishers compete fiercely in this market: Big names like Amazon, with DSSTNE, Google with TensorFlow, IBM with Watson, Microsoft with Cortana CNTK…. But also companies less known to the general public like Caffe 9a project initiated by UC Berkeley, H20 (published by the Californian company H20), Theano, (from the University of Montreal) Torch (used and improved by Facebook engineers.)
The challenge? The more developers adopt their software, the more likely they are to buy their hardware or cloud solution. Some actors (Caffe, H20, TensorFlow, Theano, Torch) have even bet on an open source diffusion: The more developers that use their engine, the more applications will be developed, so that the whole community will benefit — which will in turn increase the interest of using the software in question.
Another advantage: the more a neural network is used, the more it learns, thus becoming more effective. “Some higher level learning will come to improve our platform, but it only concerns learning that does not include any notion specific to the data of our customers,” says Jean-Francois Gagne, co-founder at PDG AI Element, a Montreal start-up.
Another way of working
But, once the AI system is in place, the most complicated, according to experts, remains the adoption of AI by all other employees. “I always warn our customers: algorithms, technologies, it’s 30% of the project; how you are going to make money is by rethinking how teams will work with AI. And that’s 70% of the work!” Says Sylvain Duranton, CEO of BCG Gamma, the global BCG AI activity.
“AI does not forget anything, goes faster, operates on a broad spectrum of knowledge and can therefore take employees out of their comfort zone by proposing tracks that they did not necessarily think of,” says Jean-Philippe Desbiolles, (IBM Watson.) In general, individuals do not like change. Coaching, good management and training are therefore essential to the process.
Copyright: Les Echos(France) / Worldcrunch – Jacques Henno.