Advancements in AI Safety: SafeAI 2021

Table of Contents

1. Introduction
2. The Importance of Trustworthy AI
3. Technological Pillar
   • Developing a New Way to Engineer AI Systems
4. Evaluation Pillar
   • Objectively Evaluating the Trustworthiness of AI Systems
5. Norms Pillar
   • Adapting and Modifying Current Standards for AI
6. Challenges in Trustworthy AI
   • Lack of Robustness
   • Unknown Factors
   • Data Obsolescence
7. Industrial Expectations
   • Collaboration with Industrial Partners
8. Scope of the Project
   • Data-Driven AI
   • Knowledge-Based AI
   • Hybrid AI
   • Distributed and Embedded AI
9. Development Environment
   • Algorithm Engineering
   • Data Engineering
   • Safety Aspects
   • Human Factors
   • System Engineering
10. Testing and Industrial Use Cases
11. Deliverables and Outputs
12. Collaboration and Expansion
13. Conclusion

Trustworthy AI: Developing a New Paradigm

Artificial Intelligence (AI) has rapidly evolved over the years, becoming an indispensable tool across various industries. However, with its increasing integration, the need for trustworthiness in AI systems has become crucial. The French government recognizes this urgency and has launched a significant project called "conference.ai" to address this concern. In this article, we will Delve into the various aspects of trustworthy AI, the challenges it presents, and the solutions proposed by the project.

Introduction

In the opening session of the "conference.ai" project, it was emphasized that trusting AI is essential, as it is expected to become an unavoidable tool for every application in the industry. Just like classical software or electronics, AI must be trustworthy to be effectively utilized. This article highlights the importance of trust in the development of AI systems, focusing on the three pillars of the "conference.ai" project: the technological pillar, the evaluation pillar, and the norms pillar.

The Importance of Trustworthy AI

Trustworthy AI is crucial for its widespread adoption and acceptance in various industries. In the Context of the "conference.ai" project, trust comes from ensuring the quality, ethics, and social impact of AI systems. To establish trust, the project emphasizes the need for certification, safety, security, and privacy data protection. However, developing trustworthy AI systems poses several challenges that need to be addressed.

Technological Pillar

The technological pillar of the "conference.ai" project aims to engineer AI systems in a new way. The classic methods of software development and engineering do not directly Apply to AI systems. Therefore, a new approach is necessary to build trustworthy AI. This involves algorithm engineering, data engineering, safety aspects, human factors, and system engineering. By considering these aspects, the project aims to develop AI systems that meet industrial expectations.

Evaluation Pillar

The evaluation pillar focuses on objectively evaluating the trustworthiness of AI systems. Traditional certification processes do not apply to AI systems, necessitating the development of new evaluation tools and methodologies. These tools will enable the assessment of the reliability, safety, and accuracy of AI components within the overall system.

Norms Pillar

The norms pillar addresses the need for adapting and modifying current standards for AI systems. Existing certification processes and norms often fail to capture the unique requirements of AI. It is necessary to redefine and Align these standards with the state-of-the-art technologies to ensure they are realistic and can be effectively implemented.

Challenges in Trustworthy AI

Trustworthy AI faces several challenges that need to be overcome. The lack of robustness, unknown factors, and data obsolescence pose significant obstacles to developing and maintaining trustworthy AI systems. Robustness is crucial, as AI systems should be able to perform in various conditions and scenarios. Unknown factors and data obsolescence refer to the changing environment and the need to continuously update AI systems with Relevant and up-to-date information.

Pros:

• Establishing trust in AI systems leads to wider adoption and acceptance.
• Addressing the challenges of robustness and data obsolescence enhances the reliability and relevancy of AI systems.
• Adapting norms and standards for AI ensures realistic and feasible implementation.

Cons:

• Developing trustworthy AI systems requires significant research and development efforts.
• Ensuring robustness and staying updated with evolving data are ongoing challenges.

Industrial Expectations

The "conference.ai" project collaborates with nine industrial partners, including companies like Renault, Dallas, Julius, Valeo Saffron, Naval Group, EDF, Atos, Early Kid, and Airbus. These partners represent various sectors, such as automotive, aeronautics, defense, energy, and manufacturing. The project aims to meet the expectations of the industry and focuses on industrial use cases to validate the effectiveness of developed AI systems.

Scope of the Project

The scope of the "conference.ai" project is extensive, covering different types of AI, including data-driven AI, knowledge-based AI, hybrid AI, and distributed and embedded AI. It acknowledges the limitations and challenges associated with each type of AI and aims to develop solutions that can be applied to real-world industrial systems.

Development Environment

The development environment for trustworthy AI requires Attention to several aspects. Algorithm engineering ensures the development of effective AI algorithms. Data engineering focuses on collecting and managing high-quality data for training and validation. Safety aspects ensure AI systems comply with safety standards. Human factors consider the interaction between AI systems and humans. System engineering ensures the integration and proper functioning of AI systems within larger systems.

Testing and Industrial Use Cases

The "conference.ai" project emphasizes testing and validation using real-world industrial use cases. The effectiveness of developed AI systems will be demonstrated through these use cases, ensuring that the project meets the expectations of the industry. Use cases from automotive, aeronautics, defense, energy, manufacturing, and other domains provide practical scenarios for testing the trustworthiness of AI systems.

Deliverables and Outputs

The "conference.ai" project aims to deliver several outputs to enhance trust in AI systems. These include validation tools for industrial use cases, taxonomies of trustworthy AI, engineering handbooks for AI development, robustness evaluation tools, certified AI components, and methodologies for verification, certification, and validation. The project aligns with industry requirements and focuses on practical implementation in real-world systems.

Collaboration and Expansion

While the "conference.ai" project initially focuses on French activities, it recognizes the need for collaboration globally. The project aims to expand its collaborations beyond France, seeking partnerships in other European countries, Canada, the United States, and other parts of the world. The objective is to Create a collaborative network that can collectively work towards developing trustworthy AI systems and establishing global standards.

Conclusion

The "conference.ai" project is a significant undertaking by the French government to address the critical need for trustworthy AI systems in various industries. By focusing on technological advancements, evaluation methodologies, and adaptations of norms, the project aims to foster trust and confidence in AI systems. The collaboration with industrial partners and the emphasis on real-world use cases ensure practical implementation and applicability. Through the project's deliverables and outputs, the development of trustworthy AI systems becomes a reality, paving the way for safe, reliable, and ethically sound AI applications in the future.

Highlights

• The "conference.ai" project addresses the urgent need for trustworthy AI systems in various industries.
• Trustworthy AI is crucial for widespread adoption and acceptance.
• The project focuses on three pillars: technological development, evaluation, and norms adaptation.
• Challenges in developing trustworthy AI include robustness, unknown factors, and data obsolescence.
• Collaboration with industrial partners and testing with real-world use cases ensure practicality and applicability.
• The project aims to deliver various outputs, including validation tools, taxonomies, and methodologies for certification and verification.
• Global collaboration and expansion are sought to establish universal standards for trustworthy AI.