Following eighteen months of intense research about the design and the structure of an automated, intelligent, energy efficient cloud/edge resources orchestrator for human-centred Artificial Intelligence (AI) system, the Horizon Europe TALON project is intensifying its integration efforts that are led by Netcompany-Intrasoft. The integration approach of the project leverages methodologies and best practices successfully applied in many other projects with complex integration requirements. Many of these practices are ground on popular agile development and DevOps methodologies. Nevertheless, the TALON orchestrator is destined to support the emerging wave of human-centred AI applications based on Industry 5.0 paradigm, which poses novel integration challenges. Following paragraphs discuss some of the most prominent of these challenges i.e., the top five integration challenges of human-centric AI systems.
1. Integration Complexity stemming from diverse components and technologies
One of the primary challenges in integrating human-centric AI systems stem from very diverse and novel technologies that they comprise, including different types of AI technologies such as edge AI technologies and decentralized federated machine learning systems. These systems often involve multiple hardware devices, software components, and AI algorithms that need to work seamlessly together. Ensuring compatibility, interoperability, and synchronization between these diverse components is a very challenging task. Integration efforts may involve integrating sensors, actuators, data analytics frameworks, and different communication protocols, which require expertise in various domains. To alleviate this complexity, we rely on the TALON modular architecture, which breaks down complex integration problems in more fine-grained and better manageable integration tasks.
2. Optimization of conflicting parameters
The TALON human-centric AI systems involve the optimization of different parameters that can be conflicting in nature. This is quite typical in cloud/edge computing systems such as the systems developed based on the TALON orchestrator, yet the challenge appears intensified in cases where human-centric AI paradigms are involved. Parameters such as latency, energy efficiency, and security need to be carefully balanced to provide optimal user experience and performance. For example, reducing latency might require processing data locally, but this could compromise energy efficiency. Similarly, enhancing security might add extra computational overhead, leading to increased latency. Striking the right balance between these parameters can be a significant challenge, requiring sophisticated algorithms and trade-off analysis. The resolution of these trade-offs is reflected in the integration and testing plans of the TALON orchestrator.
3. Testing and validation with the human in the loop
Unlike traditional software systems, human-centric AI systems require testing and validation with the human in the loop. These systems learn from human interactions and adapt their behavior accordingly. This makes it very important to involve human participants during the testing phase, which can be time-consuming and expensive, as it necessitates recruiting and coordinating a diverse group of users for testing purposes. Additionally, the involvement of humans introduces subjectivity and variability, making the testing process more challenging compared to automated testing methods commonly used in classical DevOps environments and agile software development processes.
4. Compliance with regulations and auditing requirements
Human-centric AI systems must comply with regulations and adhere to auditing requirements. For instance, the General Data Protection Regulation (GDPR) and the emerging AI Act in Europe impose strict rules on the data collection, processing, and privacy aspects of AI systems. Ensuring compliance with these regulations requires robust mechanisms for data anonymization, informed consent, and transparent data management. Moreover, strong auditing and logging mechanisms should be in place to track system behaviour and provide accountability. The latter mechanisms increase the complexity of the integration tasks, as they ask for the design and implementation of complex auditing and logging mechanisms at component, module, sub-system and system levels.
5. Ethical considerations and algorithmic bias
A fifth challenge in human-centric AI systems integration revolves around ethical considerations and algorithmic bias. AI algorithms learn from data, which can inadvertently lead to biased outcomes. Algorithmic bias might disproportionately influence certain user groups, perpetuating societal inequalities. Integrating ethical considerations into AI systems requires careful attention to data selection, algorithm design, and continuous monitoring to mitigate bias and ensure fairness. This affects in turn testing scenarios and validation practices.
Overall, the TALON team understands integrating human-centric AI systems poses novel challenges. Overcoming these challenges requires not only strong integration leadership and coordination, but also close collaboration between different TALON partners (including experts from various domains) and a holistic approach that balances technical, ethical, and legal considerations. To stay up to date with respect to the TALON Industry 5.0 systems and use cases, make sure that you follow our project in social media:
This piece has been authored by John Soldatos, Netcompany-Intrasoft.