Data science, Trustworthy AI

Our expertise in Data-Driven AI includes traditional machine learning and advanced deep learning technologies. We specialize in natural language processing (NLP) and in developing reliable and compliant Large Language Models (LLMs), We apply these techniques across various domains -such as agriculture and health- keeping the focus on the trustworthy AI aspects of linguistic technology development.

Examples include developing chatbots tailored to specific sectors and using techniques like Generative AI (GenAI) and Large Language Models (LLMs) to summarize text and answer queries. We also generate synthetic datasets to protect privacy and address data scarcity.

We excel in building semantic models like ontologies and knowledge graphs for specific domains, such as the labour market and energy distribution. These models enable us to label raw data with metadata and reason with it in an open and context-aware manner. Notable applications include creating the Semantic Interoperability Framework, adopted by multiple companies, and ensuring interoperability between IoT devices and databases.

Our Hybrid AI approach integrates data-driven and knowledge and reasoning-based methodologies, focusing on data quality and trustworthiness. By combining these approaches, we create comprehensive AI systems that provide reliable and transparent outcomes. An example is tuning LLMs for specific applications, such as a speech-based interface for in health-related questions, ensuring that AI-generated answers are valid and trustworthy.

Our commitment to Trustworthy AI ensures that our technologies conform to ethical standards and societal expectations, promoting fairness, transparency, and accountability.

This example shows how Data Science expertise in knowledge modelling and interoperability is supporting the transition towards skills-based labour market, resulting in more job opportunities for the citizens, better career mobility, reduced unemployment and more personal fulfillment.

The labor market is tight and rapidly changing, and diplomas and CVs no longer help identify the best candidate for a job. Employers and job seekers focus too much on education and experience, and not enough on skills. Better coordination between the labor market and vocational education is needed to ensure everyone finds the right job.

Understanding skills provides the best opportunities for jobseekers to adapt to the changing market, and for employers to find the perfect match for their positions.

In collaboration with UWV, SBB, CBS, and the Ministry of Social Affairs and Employment, we have applied our expertise on knowledge modelling, and our experience in connecting different stakeholders to develop CompetentNL, a common language, used to describe skills across professional and study programs. CompetentNL is developed to be a dynamic language, able to adapt to the changes of the labour market, and it is kept up-to-date to the existing European standards. Last, but definitely not least, specific attention has been dedicated to mitigating (gender) biases and provide more fair decision support.

CompetentNL aims to support the Netherlands towards a skills-based labour market, resulting in more job opportunities for the citizens, better career mobility, and reduced unemployment. The broad interest to the project, both in the public and private sector, and a number of successful pilots demonstrate the positive impact that this solution can have.

This example shows how semantic interoperability can improve citizens’ daily life and make smart energy management accessible in homes and buildings, resulting in higher energy efficiency, lower energy bills, increased awareness of energy usage, sustainable behavior, reduced power grid load, and easier connectivity of devices from different vendors.

Optimizing energy use is a complex problem, whose solution effective communication and coordination between different systems, devices, and applications.

Semantic interoperability is a technology that enables this exchange of information and interoperability. Data Science has developed key AI driven solutions: Smart Applications REFerence ontology (SAREF) – the first EU standardized IoT ontology framework - and Knowledge Engine (KE) - an ontology based open-source middleware for smart information integration across systems. Since 2014, SAREF and KE have been used in multiple B2B projects in different domains (e.g. horticulture, agrifood, smart homes, energy and defense), as well as in big pilots within the HorizonEU project InterConnect (2019-2024).

This example shows how data science supports public organisation to adopt AI solutions in a trustworthy and responsible way, resulting in better and more fair services for the citizens, and improving the trust of society in governmental bodies.

While AI is increasingly important in our society, its development and integration in existing processes poses risks and raises doubts, especially within the public sectors. The AI oversight Lab aims to support organisations and public entities to guarantee that the adoption of AI solutions happens in a responsible way. This is made possible via an interdisciplinary approach that includes both technical and organisational aspects.

The AI oversight Lab moved its first steps in 2021, when the Data Science department collaborated with the municipality of Nissewaard to evaluate biases issues and robustness of a fraud assess AI model for social benefits; several issues were identified, which led to the withdrawal of the algorithm. This first experience led to many other collaborations, for example with the Immigration and Naturalisation Service (IND) or with the Public Prosecution Service (OM). The Lab keeps growing and broadening its impact, publishing academic papers and facilitating the sharing of experiences and best practices across its community of committed stakeholders.

This example demonstrates how Data Science expertise in language technologies can lead to the creation of a sovereign and compliant Dutch Large Language Model.

The GPT-NL project aims to develop a large language model (LLM) tailored specifically for the Dutch language and culture, supportive of transparency, inclusivity, and in alignment with Dutch and European values. The project addresses challenges related to biases, ethical standards, and user data privacy. By working towards compliance to the General Data Protection Regulation (GDPR) and the EU AI Act, GPT-NL seeks to create a Dutch-language model that does not infringe on privacy or intellectual property rights. This approach aims to inspire other initiatives to innovate with generative AI while respecting legal boundaries and promoting sustainability.

The project is well-known beyond TNO, SURF and NFI. Team members have been invited as keynote speakers at Dutch conferences, and as invited speakers and panellists at both Dutch and international (European) events, and have been interviewed by organizations such as NPO, iBestuur or Computer Idee. The team is moreover continuously working on building an ecosystem, reaching out to and understanding the concerns, needs and hopes of potential data contributors, potential end users of the model, interested academics, supervisory authorities such as the Dutch privacy protection agencies, and policy makers.