Offene Abschlussarbeiten

Text-to-Image Generative Models, such as DALL-E or IMAGEN, are capable of generating images based on textual descriptions. However, these model may occasionally produce  images that do not align with the intended descriptions or display biases.

Analogous to the prompt-based analyses of Rassin et al. (2022), Bianchi et al. (2023) or Wan et al. (2024), this research aims to (1) measure the alignment between images and textual descriptions across various linguistic formalisms, or to (2) identify potential stereotypical biases in the visual content, respectively.

DrawBench provides inspiration for prompts related to cardinality and composition.

Rassin et al. (2022). DALLE-2 is Seeing Double: Flaws in Word-to-Concept Mapping in Text2Image Models.

Bianchi et al. (2023). Easily Accessible Text-to-Image Generation amplifies Demographic Stereotypes at large Scale.

Wan et al. (2024). The Male CEO and the Female Assistant: Probing Gender Biases in Text-To-Image Models Through Paired Stereotype Test.

Development of an AI Compliance Platform in Response to AI Regulations

Overview

Artificial Intelligence (AI) is transforming industries, driving innovation, and raising new ethical and regulatory challenges. As governments around the world begin to enact regulations to ensure the safe and responsible deployment of AI technologies, there is a growing need for robust mechanisms to ensure compliance. This thesis will focus on the development of an AI Compliance Platform, designed to help organizations navigate and adhere to emerging AI regulations effectively.

Objectives

The primary aim of this project is to create a comprehensive platform that facilitates the monitoring, reporting, and management of AI systems to ensure they comply with legal standards. Specific objectives include:

• Researching Global AI Regulations: Understanding and documenting the landscape of AI regulations across different jurisdictions.
• Designing Compliance Tools: Developing tools within the platform to automatically assess AI systems against compliance benchmarks.
• Risk Assessment Framework: Creating methodologies to identify and mitigate risks associated with non-compliance of AI systems.
• User Interface Design: Crafting an intuitive user interface that allows stakeholders to easily access compliance information and tools.

Methodology

The project will employ a multidisciplinary approach, combining insights from computer science, law, and ethics. It will involve:

• Literature Review: Comprehensive analysis of existing AI regulations, compliance strategies, and tools.
• Software Development: Practical application of software engineering techniques to build the platform, including programming, database management, and system integration.
• User-Centered Design: Incorporation of design thinking processes to ensure the platform is user-friendly and meets the needs of various stakeholders.

Expected Outcomes

The successful completion of this thesis will result in:

• A prototype AI Compliance Platform capable of supporting organizations in adhering to AI regulations.
• A detailed report on the feasibility and implementation of the platform, including challenges and potential future developments.
• Contributions to academic and industry discussions on AI governance and compliance.

Ideal Candidate

This topic is suited for students with a background in computer science, software engineering, AI, or related fields, who are interested in the intersection of technology and policy. Skills in programming, data analysis, and an understanding of regulatory frameworks will be advantageous.

Uncovering Bias in the Space Allocation of Text-to-Image Generative AI 

Text-to-Image Synthesis (Rombach et al, 2022) offers creative potential but may also exhibit biases in how visual space is allocated to subjects from different demographic groups. Specifically, the placement of individuals in images, particularly in terms of spatial depth, may reflect stereotypical associations.

Since psychology posits that elements in the foreground are perceived as more dominant due to their visual accessibility (Arnheim, 1974; Leder et al., 2004), this study aims to examine the spatial depth allocation in response to demographic attributes. Specifically, prompts will contain explicit reference to multiple individuals characterized by different gender, race, and religious attributes.

To measure the spatial depth, the transformers library provides off-the-shelf pipelines for depth estimation:

transformers.pipeline(task="depth-estimation", model="Intel/dpt-large")

References

Arnheim (1974). On the Nature of Photography. Critical Inquiry, 1(1), 149-161.

Leder et al. (2004). A Model of Aesthetic Appreciation and Aesthetic Judgments. British Journal of Psychology, 95(4), 489-508.

Rombach et al. (2022). High-resolution Image Synthesis with Latent Diffusion Models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (pp. 10684-10695).