Optimizing Eviden’ AI-based solutions

31 Jul 2024

By integrating the results of Confiance.ai, Eviden has strengthened its AI engineering capabilities for its Computer Vision Platform and HPC product lines. This collaboration has improved regulatory compliance, model robustness and data management, while optimizing environmental impact.

 

AI adoption challenges specific to Eviden

Within its R&D, Atos applies AI technologies to its Computer Vision Platform and HPC product lines.

The challenges are manifold:

  • Future AI regulations: preparing engineering for compliance with the AI ACT and ensuring traceability throughout the entire lifecycle of AI-based software components;
  • AI security: guarantee privacy and data confidentiality;
  • AI reliability: ensuring the robustness of AI models in production;
  • Access to data to build and test AI: overcome the scarcity of real, usable data;
  • The environmental impact of AI: measuring and optimizing the energy footprint of AI technologies throughout their lifecycle.

 

Products and services affected by the Confiance.ai program

The Eviden CVP product line from Atos aims to provide a video analytics solution for video protection systems in a variety of use cases, such as property and personal security. For example, this solution can help search for people in a camera network based on their appearance, or detect abandoned luggage in railway stations. The arrangement of cameras often implies heterogeneity in image acquisition, such as different optical properties, luminosities, viewing angles or scenes. These cameras can be installed in public spaces and must comply with local regulations, such as the RGPD and the future AI-Act of the European Union, which impose high requirements in terms of traceability, robustness and security.

In addition, Atos’ HPC products, such as predictive maintenance or supercomputer performance management tools, are highly sensitive to recommendations made to operators. The use of Confiance.ai results should guarantee the reliability of predictions produced by AI models.

 

Confiance.ai’s contribution to Eviden’ strategy and challenges

The exchanges, scientific seminars and issues presented in Confiance.ai have given Eviden access to cutting-edge knowledge and scientific advances. The reports, methodological sheets and feedback on the success or failure of experiments have provided valuable time-savings for its activities.

In addition, Confiance.ai’s work has enabled us to respond fully or partially to certain challenges:

  1. AI Act readiness (work with SHS partners, interpretability of model predictions with Xplique);
  2. Robustness, detection and correction of bias (confidence score on predictions, measurement and reduction of the impact of perturbations on images, analysis and adaptation of models to overcome the phenomenon of domain shift);
  3. Data scarcity thanks to Generative AI techniques for generating synthetic data (accelerating technological intelligence).

 

Competitive advantages developed with Confiance.ai

Confiance.ai has enabled Eviden to strengthen its knowledge and skills in AI engineering. In concrete terms, this has enabled it to anticipate the adaptations needed to comply with the AI Act, and to overcome the scarcity of real data that can be captured for use cases linked to the security of goods and people. These contributions will enable Eviden to develop key differentiators on the European market. Participation in the consortium is also a guarantee of seriousness with major French and European customers.

Challenges and conditions for success for the next phase of the Confiance.ai program

The AI field is particularly dynamic, with constantly evolving technologies. The issue of trust will shift to new techniques not covered by the initial Confiance.ai program (e.g. LLM, hybrid AI). In addition, work on Generative AI and synthetic data has been initiated, but needs to be pursued to obtain results that can be exploited in industry.

 

Example of current or future implementation/deployment

To date, the developments resulting from the Confiance.ai program are being tested in the Atos laboratory and have not yet been deployed at customer sites.

Accelerating AI Act compliance with the Confiance.ai program

The Confiance.ai program has enabled Eviden to acquire in-depth knowledge of regulations, risk analysis methodologies and standardization work. It also facilitated the construction of an engineering adaptation plan to meet regulatory obligations, and initiated this process.