Periodic Reporting for period 1 - ULTIMATE (mUlti-Level Trustworthiness to IMprove the Adoption of hybrid arTificial intelligencE)
Okres sprawozdawczy: 2022-10-01 do 2024-03-31
Artificial Intelligence (AI) has entered the business mainstream, opening opportunities to boost productivity and innovation, but currently suffers limitations hindering wider adoption in industrial settings. The paradigm of hybrid AI, which complements data-driven AI approaches with physics-driven insights, has shown particular promise. However, hybrid AI does not fully address the issue of trustworthiness (validity, explainability and ethics). ULTIMATE (www.ultimate-project.eu) will pioneer the development of industrial-grade hybrid AI based on 3 stages to ensure trustworthiness and promote the widespread adoption of hybrid AI in industry. The main project objectives are the following:
- Develop data representation and visualization models for hybrid AI algorithms, and propose innovative architectures for the construction and training of trustworthy hybrid AI algorithms,
- Design evaluation methodologies for hybrid AI algorithms,
- Implement the developed hybrid AI algorithms in industrial environments and assess their performance,
- Ensure the ethical compliance and trustworthiness of the developed hybrid AI algorithms and the wide uptake of the project results at EU and international level.
The project expected impacts are:
- Advance the current knowledge on the design, development, and deployment of production-grade hybrid AI and on rigorous evaluation methodologies (e.g. confidence estimation methods) to significantly increase trustworthiness.
- Go beyond some existing standards as a reference in AI solutions to meet industrial requirements (related to safety for instance) to cover AI systems trustworthiness more adequately including social and ethical issues.
- Ensure that AI development and implementation is human-centric and is a force for good in society whilst evaluating the consequences taking into account the criteria of people (compliance with appropriate legal, ethical and societal foundations) and the machine’s criteria.
- Support the creation of high quality jobs where humans are making informed decisions using AI outputs rather than simply executing tasks they do not understand.
- Develop data representation and visualization models for hybrid AI algorithms, and propose innovative architectures for the construction and training of trustworthy hybrid AI algorithms,
- Design evaluation methodologies for hybrid AI algorithms,
- Implement the developed hybrid AI algorithms in industrial environments and assess their performance,
- Ensure the ethical compliance and trustworthiness of the developed hybrid AI algorithms and the wide uptake of the project results at EU and international level.
The project expected impacts are:
- Advance the current knowledge on the design, development, and deployment of production-grade hybrid AI and on rigorous evaluation methodologies (e.g. confidence estimation methods) to significantly increase trustworthiness.
- Go beyond some existing standards as a reference in AI solutions to meet industrial requirements (related to safety for instance) to cover AI systems trustworthiness more adequately including social and ethical issues.
- Ensure that AI development and implementation is human-centric and is a force for good in society whilst evaluating the consequences taking into account the criteria of people (compliance with appropriate legal, ethical and societal foundations) and the machine’s criteria.
- Support the creation of high quality jobs where humans are making informed decisions using AI outputs rather than simply executing tasks they do not understand.
The ULTIMATE project has seen significant strides in implementing hybrid AI algorithms in industrial environments and assessing their performance across various use cases. Key achievements and work performed include:
1. the identification of Key Performance Indicators (KPI): the project has identified specific KPI tailored to each use case and the hybrid AI algorithms themselves. These KPI serve as quantitative and qualitative measures to evaluate the efficiency and effectiveness of the algorithms in real-world industrial scenarios.
2. the development of a data visualisation toolkit for representing high-dimensional information: it enables to visualise a number of attributes from the datasets, including the statistical models that highlight the presence and absence of anomalies when mapped against time series. The use of visualisation toolkit for monitoring the network evolution and the training of deep-learning models also offers insight into the quality of the training data.
3. the development and implementation of some hybrid AI algorithms: The project has developed advanced hybrid AI algorithms tailored to specific industrial use cases, including space applications, robotic workshops, and industrial settings. These algorithms leverage a combination of Machine Learning (ML) and traditional AI techniques to address complex challenges in each domain.
Three different categories of hybrid AI algorithms are currently developed:
- a Physics-Informed ML approach to detect and classify anomalies on satellites to alleviate the human burden at ground
- a neuro-symbolic one that combines Deep Learning (DL) with symbolic knowledge for reasoning in front of potential safety issues due to a shared space between humans and robots
- a last one based on DL and Markov decision processes which aims at coordinating robots that sort and bring different types of objects within an industrial logistic chain.
1. the identification of Key Performance Indicators (KPI): the project has identified specific KPI tailored to each use case and the hybrid AI algorithms themselves. These KPI serve as quantitative and qualitative measures to evaluate the efficiency and effectiveness of the algorithms in real-world industrial scenarios.
2. the development of a data visualisation toolkit for representing high-dimensional information: it enables to visualise a number of attributes from the datasets, including the statistical models that highlight the presence and absence of anomalies when mapped against time series. The use of visualisation toolkit for monitoring the network evolution and the training of deep-learning models also offers insight into the quality of the training data.
3. the development and implementation of some hybrid AI algorithms: The project has developed advanced hybrid AI algorithms tailored to specific industrial use cases, including space applications, robotic workshops, and industrial settings. These algorithms leverage a combination of Machine Learning (ML) and traditional AI techniques to address complex challenges in each domain.
Three different categories of hybrid AI algorithms are currently developed:
- a Physics-Informed ML approach to detect and classify anomalies on satellites to alleviate the human burden at ground
- a neuro-symbolic one that combines Deep Learning (DL) with symbolic knowledge for reasoning in front of potential safety issues due to a shared space between humans and robots
- a last one based on DL and Markov decision processes which aims at coordinating robots that sort and bring different types of objects within an industrial logistic chain.
As the project just implemented the first version of the hybrid AI models (not yet evaluated), it is still premature to get results beyond the state of the art.