scieee AI-readable full text Open interactive document viewer

Specifying Multiple Paradigm and Component AI-based Systems: A Perspective Approach

Anonymous, Anonymous

Publisher
Zenodo
DOI
10.5281/zenodo.17426500
Repository
zenodo.org
Language
en

Abstract

Current AI applications are increasingly complex, often involving two or more interacting models, a trend driven by mass data production and advances in fields like computer vision and LLMs. Unlike traditional software, AI-based systems are probabilistic in nature, introducing unpredictability, and their "black box" characteristics make explaining and mitigating problematic behavior difficult, posing new challenges for software engineering and demanding new approaches to requirements engineering (RE). Despite RE being a fundamental development step, a significant gap remains in the literature for RE for AI-based software, particularly for systems with multiple AI components and paradigm, as their behavior can only be evaluated after extensive experiments, training, and testing. To address this challenge, we propose and preliminarily validate a new approach for modeling the requirements of multiple AI paradigm and components software. This approach is based on the concept of perspectives and concerns that can have associations, requiring a common core of concerns that must be specialized for different AI paradigms and capable of accounting for non-trainable models. We preliminary validated our approach by specifying one use case containing more than one AI paradigms and components and discussed through a focus group with AI academic practitioners. Results indicate that our approach are suitable for specifying systems with multiple AI paradigms and components.

Full text

Cores principais Cores secundárias 1 Specifying Multiple Paradigm and Component AI-based Systems: A case study Cores principais Cores secundárias 2 The Complex and Evolving Landscape of AI Systems AI-based systems are becoming inherently more complex, often handling more than one AI model and more than one paradigm associated with them. This growth in complexity is a direct result of advances in the AI field, which are closely linked to hardware advancements, such as microchips and video cards. These advancements are unlocking new AI capabilities that enable the solution of more intricate problems, including: ●Autonomous navigation. ●Video generation. ●Image manipulation, among many others Cores principais Cores secundárias 3 The Gap in Current Software Development Practices ●Despite the technological progress, there is a significant failure rate observed in the domain: it is estimated that numerous AI projects fail to be deployed and never progress as far as production. ●This failure can be partially attributed to the scientific community's tendency to focus primarily on model development, often neglecting the early stages of software development. ●Key challenges arising from this negligence include: ○Limitation of Traditional SE Approaches: As the AI area evolves in model capabilities and problem complexity, the more traditional approaches in the software engineering field are limited when handling the new challenges that AI introduces. ○Probabilistic Nature: This limitation exists mainly because of the nature of AI-based systems, which possess a probabilistic behavior based on data. ○Industry Need: The industry currently lacks reliable and robust SE solutions for the development of complex AI systems. Cores principais Cores secundárias 4 Our Proposed Solution and Methodology ●Our research proposes an approach specifically designed to support the Requirements Engineering (RE) step of complex AI-based systems. The goal is to allow for the modeling of AI solutions with multiple paradigms and components. ●We utilized the PerSpecML approach as a foundation. PerSpecML provides a solution for RE of ML-based systems by accounting for five main perspectives: 1. Systems objective. 2. User experience. 3. Infrastructure. 4. Model. 5. Data. Cores principais Cores secundárias Methodology Cores principais Cores secundárias 6 A Requirements Engineering Approach for Complex AI Systems ●The Problem Addressed: The complexity of modern intelligent systems has led to a lack of appropriate approaches to support the Requirements Engineering (RE) process. ○Our Solution: OmniSpecAI / PerSpecOmniAi. ●This approach focuses on supporting the RE step of complex AI-based systems, allowing the modeling of solutions with multiple paradigms and components. ○Foundation: We utilized the PerSpecML approach as our base. ●PerSpecML is an RE approach for ML-based systems defined by 5 main perspectives and requiring a multi-disciplinary team Cores principais Cores secundárias 7 The Core Structure and Relationships ●OmniSpecAI: Perspectives as System Components ●Core Idea: OmniSpecAI defines the relationship between perspective components and paradigms. The system starts from abstract, general AI concerns and specializes depending on the objective to be fulfilled. ●The 5 Perspectives: 1. System Objective Perspective (SOP). 2. User Experience Perspective (UXP). 3. Infrastructure Perspective (IP). 4. AI Perspective (IAP, previously Model Perspective). 5. Data Perspective (DP). ●Modeling Complex Systems: ○The perspectives represent components of the system (e.g., the IAP consumes data specified by DP, served on an IP, which contributes to SOP). ○Our approach allows for multiple instances of the same perspective to fulfill the goals of a single AI system. ○Component Sharing: By allowing relations between perspectives, we facilitate component sharing, such as when two distinct AI components share the same infrastructure or the same data source. Cores principais Cores secundárias 8 OmniSpecAi Cores principais Cores secundárias 9 Perspective Deep Dive (Part I) ●System Objective Perspective (SOP) ○Main Goals/ Concerns ■Represents organizational concerns: context, needs, organizational goals, and success indicators. ○Key Relationships ■Has a direct relationship with IP (1 to N infrastructure components) and UXP.