Flexible and trustworthy AI solutions for companies

Currently available AI systems are usually only suitable for solving specific tasks such as speech or object recognition. However, small and medium-sized enterprises (SMEs) need AI solutions that are specifically tailored to their problems and with which they can react flexibly to market changes. In addition, it is often difficult to understand which factors influence the decisions made by AI systems. This lack of transparency undermines trust in the technology and, in addition to the frequent lack of expertise, stands in the way of wider use of AI solutions in SMEs. In the Trust AI project, we are tackling these challenges and want to provide SMEs with easy access to AI services. An AI platform for different problems . An AI platform for different problems The aim of the Trust AI project is to set up an easily accessible platform for interactively trainable AI services that can be adapted to different problems. To achieve this, it is necessary to familiarise ourselves with the needs of future users of the platform - cultural and public institutions as well as SMEs. For this reason, we have carried out requirements analyses to identify SME-specific challenges and use cases. Based on this, we are developing concepts for personalised, interactively trainable AI solutions whose decisions are explainable and easy to understand. In the first part of the Trust AI project, we will work on three complementary digitalisation use cases: 1) Agriculture - pest and disease detection in viticulture, 2) Health - AI-supported diagnosis recommendation in clinical gait analysis and 3) Cultural heritage - visual analysis of historical manuscripts. In the future, the Trust AI platform and methodologies developed will be extended to other real-world use cases of SMEs. An innovative AI learning paradigm. The centrepiece of the Trust AI project is an innovative methodology that makes it possible to train AI models interactively and transparently. We are moving away from a rigid ‘Artificial Intelligence’ and towards a collaborative, adaptive and trustworthy ‘Assistive Intelligence’. Normally, complex AI models require large amounts of annotated data. With the Trust AI approach, however, training can also be started with a small amount of annotated data. The AI system identifies complex data samples from the non-annotated data and attempts to predict them. It then informs the user of the predictions for this data, including an explanation of why they were made. The users can then interactively reject or confirm the factors in the explanation or introduce new concepts on the basis of which the AI system should make its decision. Trust AI thus goes beyond the usual approach of only training AI models based on existing data annotations, as in our approach humans and AI are in a continuous dialogue. This reduces the dependency on annotated training data and enables flexible solutions that can be customised to specific company problems. Furthermore, the interactive dialogue between the AI model and the user allows the understanding of the task to be deepened step by step and the training progress of the model to be permanently monitored. This makes it easier to identify systematic errors in the model, minimise them and thus strengthen confidence in the AI. Figure 1: One of the three initial use cases is the AI-supported diagnosis recommendation in clinical gait analysis. In clinical gait analysis, movement and force measurement data are used for the diagnosis of gait disorders. The AI attempts to identify characteristic features for certain gait disorders in the measurement signals. The user (a clinical expert) can use the interactively explainable AI developed in Trust AI to identify the correct features and point out potentially relevant signal sequences to make it easier for the AI model to learn and avoid systematic misbehaviour. With Trust AI, we want to offer a platform that provides explainable and trustworthy AI models that require no prior technical knowledge and are accessible via easy-to-use user interfaces. We are happy to receive enquiries from companies about using our Trust AI platform and our learning methodology and look forward to working together to develop solutions for specific and individual problems (contact: matthias.zeppelzauer@fhstp.ac.at