HomeEducationCollective Learning In Artificial Intelligence AI

Related Articles

Collective Learning In Artificial Intelligence AI

Federate learning data from various organizations, then apply machine learning models to them to create robust AI that a single actor could never have developed alone. This is the advantage of collective learning.

What is collective learningin AI?

By collective learning we mean the pooling of data by several individuals or organizations. This concept, which dates back to prehistoric times, is coming back to the fore with the emergence of artificial intelligence . Why ? Because collective learning, by pooling AI training data between several organizations, makes it possible to create very robust machine learning applications. Learning models that one of the actors involved in the process could never have created alone based solely on his own volume of data, the latter obviously being less massive and, de facto, composed of less varied information.

Collective learning applied to AI is gaining momentum in Europe as a means to thwart the American digital giants. With markets which remain specific to each of its 27 member countries despite a recognized desire for homogenization, the EU does not allow the emergence of a company capable of reaching the critical mass and the volume of knowledge necessary to face their challenges. IA. The Gafam and other Natu ( for Netflix , Airbnb, Tesla  and Uber) aggregate the information of billions of users: digital journeys, purchasing preferences, geolocation… They thus constitute gigantic machine learning data sets that allow them to build robust AIs covering their key issues: product recommendation, promotional targeting, optimization of manufacturing, logistics, prices, etc. defining their R&D strategies using predictive models. In Europe, collective learning is therefore considered to be one of the only solutions capable of creating a competitive force on this side of the Atlantic.

What is the theory of collective learning?

Collective learning began to be conceptualized in the 2000s. For its first theorists, including Maarten de Laat and Robert Simons (who broached the subject in 2002) and Georges Wildemeersch (in 2007), this notion refers to the way in which a group diverse individuals work on processes of shared problematization and a shared conception of meaning and knowledge. 

According to political scientist Sarah J. Whatmore (in an article published in 2009), the theories of collective learning reflect the way in which these processes are structured and manage the complexity of a social system, its internal and external demands, and how they bring out a collective dynamic capable of learning and transforming itself in a dynamic way. 

What is collective learning?

Collective learning accompanies the appearance of Man at 200,000 years old in Africa. Since then, it is his talent for collective learning, via the preservation of information and know-how, their sharing and transmission to subsequent generations, which has enabled Man to create entirely new forms of complexity and to develop new forms of complexity. richer and richer civilizations. 

Collective learning in AI, how does it work?

“By increasing the volume of relevant learning data, collective learning contributes to reducing the feature engineering phase of machine learning”, explains Didier Gaultier, director of data science & IA at ESN Business & Decision, a subsidiary of Orange. As its name suggests, this step consists in formalizing the attributes (features) common to the learning information. “When we lack data, we increase the number of variables, or even the number of dimensions in the training base, to compensate. This de facto takes more time”, explains the data scientist ,

Didier Gaultier warns: “However, collective learning does not solve the problem of personal data or that of biases which, let us remember, are most often linked to the training base.” If there is a bias that leads to errors or injustices in the results, who would be responsible? The publisher of the model or all or part of the companies that have shared their data? Unlike ad targeting, the question will prove to be the most critical in the case of a credit granting AI or an autonomous car algorithm.

What is the advantage of collective learning?

Collective learning offers several advantages recognized in the field of artificial intelligence. “By increasing the volume of relevant data available for training AIs, it increases the complexity of the learning model and its robustness (that is to say its ability to generalize to new information, editor’s note) The accuracy of the results is mechanically optimized “, summarizes Didier Gaultier at Business & Decision.

The Californian publisher Moveworks measured the contribution of collective learning before and after its implementation in its flagship application, an IT support assistant. “By using the interaction data of a single client company, we achieve an accuracy (or rate of correct answers, editor’s note) of 30%, even with the most modern NLP models. Collective learning makes it possible to raise this figure at 60% “, congratulate Jiang Chen and Yi Liu, respectively vice-president of machine learning and director of research of the start-up in San Francisco. And that’s not all. Downstream, transfer learning is also implemented by Moveworks to optimize the processing of terminologies specific to each organization. The principle ? Integrate the best open source neural networks tailored to manage exchanges in natural language into the platform, then add additional layers to them according to the target technical vocabulary. A method which, in the end, allows Moveworks to record a level of precision of 90%.

 

Accuracy rate recorded by Moveworks following the implementation of collective learning and then transfer learning to build its NLP platform
Accuracy rate recorded by Moveworks following the implementation of collective learning and then transfer learning to build its NLP platform

Is there a European example of collective learning?

n Europe, Gaia-X is the first collective learning initiative aimed at offering a credible alternative to the American digital giants. With the aim of achieving a sovereign ecosystem of integrated cloud offers, this consortium, which currently has 212 members , also sets itself the mission of propelling sectoral platforms for federating data. Based on a standardized architecture (see diagram), they must boost synergies between fields of activity and, ultimately, promote the emergence of new services which can obviously be backed by AI. It is therefore a large-scale collective learning approach.

Seduced by the approach, French people from several sectors joined Gaia-X, each supported by one or more players. This is the case in finance and insurance (with the Caisse des Dépôts et Consignation), in energy (EDF), in mobility (Amadeus and Air France KLM), in space and satellite data (Dassault Systems, EBRC), in aerospace (Airbus, Thales Alenia Space), in the green (Engie), in agriculture (General Association of Wheat Producers) or even in health (with the Health Data Hub). 

Architecture of the sectoral data spaces of the Gaya-X project. © Gaya-X
Architecture of the sectoral data spaces of the Gaya-X project. © Gaya-X

“Collective learning is not new. E-merchants in the North of France, foremost among them La Redoute and the 3 Suisses, have been sharing marketing data for years with a view to feeding their algorithms.targeting “, comments Didier Gaultier. Founded in 2006 by Didier Farge, serial entrepreneur and ex-DoubleClick, the French company Conexance very quickly felt the vein. This Lille company offers e-merchants to share their digital data anonymously in in order to optimize their cross-selling and up-selling actions and, more generally, product recommendations. The underlying solution is unsurprisingly built around a predictive machine learning engine. Ad retargeting Criteo follows the same logic. Bringing together data from tens of thousands of companies, the group of French origin, now listed on the Nasdaq, opened in Paris in 2018 a research laboratory on applied AI to advertising, with a budget of 20 million euros.

What could be the future of collective learning?

Public laboratories are working on what the future of collective learning could be. Among the main luminaries in the field, the Chilean César Hidalgo, former director of the Collective Learning group of the MIT Media Lab, has headed the Augmented Society chair of the Institute of Artificial and Natural Intelligence at the University of Toulouse since 2019. His approach: apply machine learning and statistics to masses of data shared in open data to better understand macroeconomic movements.

“Artificial intelligence is limited for the time being to specific tasks (recognition and speech synthesis, texts, images, sentiment analysis … editor’s note). In the future, the appearance of generalist AI capable of carrying out multiple tasks will turn things upside down, “said César Hidalgo. “If countries and companies agree to share these next-generation models, we will see the emergence of multicellular AI networks capable of understanding and solving problems much more difficult to deal with, involving multiple events and interactions in real time. ” Such a network could make it possible to grasp dynamic phenomena based on complex systems, such as financial flows, climate change, population movements,

What are the synonyms of collective learning?

Collective learning has many synonyms: combined learning, joint learning, cooperative learning, joint learning, shared learning or even unified learning.

Latest Posts

error: Content is protected !!