WP 4 will develop proof-of-concept applications to showcase the benefits of our frameworks focusing on PEP principles investigated in WP 1,2 and 3. CUHK will especially focus on speech translation technologies and medical applications within the Neural Mesh context. Within these areas, team members from CUHK already have available to them a self-developed LLM and a growing ecosystem of supporting in-house technologies for speech translation, speech-based rehabilitation and other health-related applications. Technological acceptance of these solutions will also be investigated, allowing for a credible case to be made to potential investors following the project, thereby helping to integrate the results into commercial products.

To demonstrate the results of WP 1,2 and 3, we plan to provide solutions to benchmark problems of increasing complexity, as well as develop proof-of-concept applications within the fields of speech translation and health/medicine. Among the benchmarks we will consider, the simplest ones include those that have been used for decades, such as the MNIST, CIFAR, the California Housing, and Wine Quality datasets. The more complex benchmarks are the datasets widely used today, like ImageNet, SuperGLUE, and MMLU. The simpler benchmarks will be valuable for gaining deeper insights into the benefits and potential limitations of our frameworks, as they yield interpretable results and are already well understood. This will help us iterate on our methods more efficiently. On the other hand, evaluating on complex benchmark problems will allow us to highlight the advantageous properties of our results in terms of the three aspects of precision, explainability and provability.

Besides the above benchmarks, we will also develop different applications to showcase our results. CUHK will focus on both speech technology applications and medical applications in the Hungarian context. Within these domains, team members at CUHK already have years of experience with and a deep technical know-how in developing speech-based diagnostics tools (e.g., for dysarthria, or neurocognitive disorders such as Alzheimer’s Disease), as well as in developing language models in well-defined areas such as legal or financial domains. Within this project, CUHK will apply their expertise combined with the different PEP AI frameworks to showcase the Precision, Explainability and Provability related aspects of custom-developed speech-based diagnosis and translation models. In these education, speech and medical domains, high precision reflects the ability of the system to provide only information that is objectively and factually correct, while explainability revolves around its ability to “explain its work” in terms of sources and other implicit notions that were instrumental in producing the result. Finally, provability in this context refers to our ability to prove that the behavior of the system will not deviate in any meaningful way from the scope for which it was developed. All three criteria will be thoroughly investigated with respect to the proof-of-concept applications.

The factors determining the technology acceptance of PEP AI solutions can be detected using marketing science models (e.g., TAM and extended versions), quantitative sampling (e.g., CAWI), and covariance-based structural equation modeling (e.g., CB-SEM). In addition, the conceptual description of a fine-tuned AI maturity model will also be a novel feature of current project. With this, the social perceptions of the technology, and the criteria for successful technology-acceptance and diffusion will be determined. The project will also address the application of the PEP AI principles in the context of the EU AI Act.