Discovering Insights in AI Safety-Critical Systems: Poster Highlights

Table of Contents:

1. Introduction
2. Poster 1: Combining Data-Driven and Knowledge-Based ai Paradigms for Engineering AI Safety-Critical Systems
3. Poster 2: Out of Distribution Detection Based on Clustering in the Embedding Space
4. Poster 3: Safety Concerns and Mitigation Methods for Visual Deep Learning Algorithms
5. Poster 4: Framework to Argue Quantitative Safety Targets in Assurance Cases for AI/ML Components
6. Poster 5: The Dilemma Between Data Transformations and Adversary Robustness
7. Poster 6: Leveraging Vision Transformers to Increase the Robustness of Safety-Critical Systems
8. Conclusion
9. Resources
10. FAQ

👉 Introduction

In this article, we will explore the exciting field of AI safety-critical systems. We will dive into various topics discussed in six posters presented at the conference.ai. The posters cover a range of themes, including combining data-driven and knowledge-based AI paradigms, out of distribution detection, safety concerns and mitigation methods for visual deep learning algorithms, quantitative safety targets in assurance cases, the dilemma between data transformations and adversary robustness, and leveraging vision transformers to enhance the robustness of safety-critical systems.

Now, let's have a closer look at each poster and delve into the valuable insights presented by the authors.

👉 Poster 1: Combining Data-Driven and Knowledge-Based AI Paradigms for Engineering AI Safety-Critical Systems

The first poster tackles the challenge of deploying AI in safety-critical systems. It emphasizes the need to demonstrate the correctness, validity, interpretability, and robustness of AI-based systems. The authors propose a transfer CI engineering analysis framework that formalizes all engineering activities and roles. They introduce an end-to-end engineering workbench that combines data-driven AI and knowledge-based AI paradigms. The workbench aims to assess risk, demonstrate safety properties, and provide trustworthy characteristics and key performance indicators (KPIs) for the sound deployment of AI interpretation systems.

👉 Poster 2: Out of Distribution Detection Based on Clustering in the Embedding Space

The Second poster focuses on detecting out-of-distribution samples using clustering in the embedding space. The authors explore the use of contrastive learning to cluster similar instances together and push apart dissimilar instances. They compare Supervised and self-supervised methods and evaluate the quality of cluster formations using metrics such as global separation and cluster purity. The results show interesting trends in the performance of different methods and provide insights into the effectiveness of clustering for out of distribution detection.

👉 Poster 3: Safety Concerns and Mitigation Methods for Visual Deep Learning Algorithms

Poster 3 provides an overview of safety concerns and mitigation methods for visual deep learning algorithms. The authors categorize the faults and underlying causes that can arise in these algorithms. They discuss the importance of considering data sets, architecture, testing, and the inherent nature of machine learning in ensuring safety. The poster highlights the limitations of existing mitigation methods and suggests the need for task-specific standards to address safety concerns effectively.

👉 Poster 4: Framework to Argue Quantitative Safety Targets in Assurance Cases for AI/ML Components

The fourth poster introduces a framework to argue for quantitative safety targets in assurance cases for AI/ML components. The authors propose a two-line argument approach to demonstrate the reduction of safety risk and the satisfaction of quantitative goals. They emphasize the importance of considering uncertainties and runtime predictions to detect high-risk situations. The poster emphasizes the need for structured assurance cases tailored to AI-enabled systems and provides insights into building a persuasive argument for safety.

👉 Poster 5: The Dilemma Between Data Transformations and Adversary Robustness

Poster 5 explores the trade-off between data transformations and adversary robustness in unintentionally introducing vulnerabilities in deployed models. The authors focus on dimensionality reduction, feature selection, and trend extraction techniques. They analyze the impact of these transformations on the resulting data manifold and highlight the importance of considering a dataset's intrinsic characteristics to avoid vulnerabilities. The poster provides a comprehensive understanding of the relationship between data transformations and robustness.

👉 Poster 6: Leveraging Vision Transformers to Increase the Robustness of Safety-Critical Systems

The last poster presents the application of vision transformers to enhance the robustness of safety-critical systems. The authors compare vision transformers with convolutional nets and demonstrate how an ensemble of these models can improve common corruption robustness. The poster highlights the benefits of using vision transformers and suggests avenues for further research, such as exploring object detection and Image Segmentation problems.

👉 Conclusion

In conclusion, the posters presented at conference.ai shed light on various aspects of AI safety-critical systems. The authors provide valuable insights into combining data-driven and knowledge-based AI paradigms, out of distribution detection, safety concerns and mitigation methods, quantitative safety targets, the impact of data transformations, and leveraging vision transformers. These findings contribute to the ongoing efforts in ensuring the safety and reliability of AI systems in critical domains.

👉 Resources

• Link to Conference.ai: conference.ai
• Link to Poster Exhibition: poster exhibition

👉 FAQ

Q: What is the main focus of the conference.ai poster presentations? A: The conference.ai poster presentations cover a wide range of topics related to AI safety-critical systems, including combining AI paradigms, robustness, safety concerns, quantitative targets, data transformations, and leveraging vision transformers.

Q: How can data transformations impact the robustness of AI models? A: Data transformations can introduce vulnerabilities in AI models, leading to decreased robustness. It is crucial to understand the intrinsic characteristics of the dataset and carefully consider the impact of transformations such as dimensionality reduction, feature selection, and trend extraction to maintain robustness.

Q: What is the significance of out-of-distribution detection using clustering in the embedding space? A: Out-of-distribution detection is essential in ensuring the reliability and safety of AI systems. Clustering in the embedding space can help distinguish between in-distribution and out-of-distribution samples, thereby enhancing the robustness of AI models and increasing their ability to handle diverse data.

Q: How can assurance cases be used to argue for quantitative safety targets in AI/ML components? A: Assurance cases provide a structured approach to argue for quantitative safety targets in AI/ML components. By considering uncertainties, runtime predictions, and reducing safety risks through comprehensive engineering activities, assurance cases can establish a persuasive argument for the safety and reliability of AI systems.

Q: What is the impact of vision transformers on the robustness of safety-critical systems? A: Vision transformers offer potential improvements in the robustness of safety-critical systems. By leveraging ensemble methods and comparing vision transformers with convolutional nets, it is possible to enhance common corruption robustness. Further research is needed to explore the application of vision transformers in other tasks such as object detection and image segmentation.

Q: Where can I find more information about the posters presented at conference.ai? A: You can find more information about the posters presented at conference.ai in the poster exhibition section of the conference website. Visit poster exhibition for detailed insights into each poster's content.