Ihmisen ohjaama yhteistyö, monitavoite, selitettävissä,
oikeudenmukainen ja yksityisyyttä suojeleva tekoäly digitaaliseen terveyteen (HarmonicAI)

Päärahoittaja

Euroopan komissio

Rahoittajan antama koodi/diaarinumero: 101131117

Päärahoittajan myöntämä tuki (€)

101 200,00

Rahoitusohjelma

MSCA muut, HE (2021-2027) (MSCA Marie Skłodowska-Curie Ac...)

Hankkeen aikataulu

Hankkeen aloituspäivämäärä: 01.01.2024

Hankkeen päättymispäivämäärä: 31.12.2027

Tiivistelmä

Innovative AI models, algorithms, and techniques are being proposed and developed by HarmonicAI- project. However, the current research efforts to tackle the three challenges are fragmented and have culminated in a variety of solutions with heterogeneous, non-interoperable,
or even conflicting capabilities. Explainability, fairness and privacy are
inherently intertwined with a complex conflicting and/or complementary
relationship between any two of them. Achieving explainability normally
demands a sufficient amount of data for feature extraction and feature
engineering, which is against the initiative objective of reducing data
collection in AI privacy research. Similarly, data collection across a wider
spectrum of populations will help mitigate bias arising from representative
data. However, it has a destructive impact from the privacy perspective. Both
explainability and fairness benefit from a more inclusive dataset in principle. But an intrinsic or post-hoc explainable model is usually trained using a pre-defined selection of features. Neglecting implicit hidden features could result in damaging biases. Moreover, bias mitigation techniques that reformulate the AI models by incorporaning discriminatory behaviours into the objective function complicates the design of explanation methods, making the developed fair model more difficult to explain. In the light of the complex intertwining relations across explainability, fairness and privacy, fragmented research in each individual area without interacting with one another is not a
sustainable pathway to achieve trustworthy AI. To this end, HarmonicAI is therefore aimed at building a multi-objective co-design methodology which can take the three conflicting and/or complementary objectives together into account to develop a coherent AI system that meets individual requirements, and at the same time, settles on an agreed trade-off.
The realisation of trustworthy AI must be built in ways that ensure humans are always in ultimate control and responsible for what the AI system will do. This is particularly crucial for digital health applications which involves decisions that affect a person’s life, health and quality of life. HarmonicAI will adopt a human-in-the-loop design approach, which involves a variety of healthcare stakeholders throughout the lifecycle of design, development, and evaluation. The research and innovation will be centered around the real-world sector
requirements gathered from relevant stakeholders. Human experts’ domain knowledge and analytical reasoning will guide the multi-objective design of explainable, fair and private AI models. This is both to ensure that the AI decisions are made carefully under human supervision, and to maintain the role of AI systems in supporting humans.