ABSTRACT. Generative AI, particularly large language models, is expected to bring about substantial change at the level of users, organizations, and society. While users over the last decade or so have become familiar with conversational interactions with computers though use of chatbots, the availability of large language models – and the benefits of text and image generation through large language models or text-to-image services – have opened a wide range of new use cases. Individual users have already taken up generative AI, particularly for productivity purposes. Organizations regard generative AI as holding high potential benefit but also to entail important challenges, e.g., in terms of security and privacy, as well as new forms of competition. At the level of society, generative AI has been the focus of substantial public debate. While there is increasing agreement on the need for regulation and means to control the development of generative AI, this may be more efficient if it seeks to guide its development rather than curb it. Through human-oriented technology research, we can help guide the impact of generative AI to the benefit of users, organizations, and society. In this talk, I will discuss the knowledge base for such a human-oriented approach and point out important future research needs.

ABSTRACT. Accessing previous data when updating the model with new data is a common problem in some incremental learning applications. This prevents, for example, neural networks from suffering catastrophic forgetting. In this paper, we focus on the incrementing of NCMFs (Nearest Class Mean Forests) for which access to old data is required with classical incrementing strategies such as IGT (Incremental Growing Tree). We propose a new incrementing strategy, named IGTLGSS (Incremental Growing Tree with Local Generation of Synthetic Samples), that allows these kind of random forests to continue to increment without relying on old data. For this purpose, the old data are replaced by synthetic data that are generated from the pre-trained NCMF which has to be incremented. Experimental studies are performed on UCI benchmarks. The results show that, for the used datasets, NCMFs are able to generate realistic synthetic data. Moreover, the first results obtained following the assessment of our incrementing strategy are encouraging.

ABSTRACT. Detecting outliers or anomalies is a common data analysis task. As a sub-field of unsupervised machine learning, a large variety of approaches exist, but the vast majority treats the input features as independent and often fails to recognize even simple (linear) relationships in the input feature space. Hence, we introduce RECol, a generic data pre-processing approach to generate additional columns (features) in a leave-one-out fashion: For each column, we try to predict its values based on the other columns, generating reconstruction error columns. We run experiments across a large variety of common baseline approaches and benchmark datasets with and without our RECol pre-processing method. From the more than 88k experiments, we conclude that the generated reconstruction error feature space generally seems to support common outlier detection methods and often considerably improves their ROC-AUC and PR-AUC values. Further, we provide parameter recommendations, such as starting with a simple squared error based random forest regression to generate RECols for new practical use-cases.

ABSTRACT. Colorectal carcinoma is a leading cause of mortality worldwide and predominantly originates from colon polyps. Not all polyps metamorphose into carcinomas, therefore polyps are categorized via various classification systems. The advent of deep learning and the proliferation of video data have given rise to a plethora of model architectures for automated video classification of polyps. However, the selection of an appropriate model for specific tasks requires careful consideration of various factors, including performance metrics and computational efficiency. In this paper, we present a comparative study of six state-of-the-art model architectures. Capitalizing on the strengths of several state-of-the-art models, a newly developed voting system enhances classification accuracy while maintaining computational efficiency, demonstrated across multiple distinct datasets. The paper explores the integration of such a voting system within the broader framework of video-based polyp identification and provides an empirical evaluation of its performance when juxtaposed with contemporary models. The findings underscore the potential of the proposed system in advancing colorectal polyp classification methodologies, aiming to contribute to early and accurate polyp classification, which is vital in preventing colorectal cancer.

ABSTRACT. Matrix production systems (MPSs) offer significant advantages in flexibility and scalability when compared to conventional line-based production systems. However, they also pose major challenges when it comes to finding optimal decision policies for production planning and control, which is crucial to ensure that flexibility does not come at the cost of productivity. While standard planning methods such as decision rules or metaheuristics suffer from low solution quality and long computation times as problem complexity increases, search methods such as Monte Carlo Tree Search (MCTS) with Reinforcement Learning (RL) have proven powerful in optimising otherwise inhibitively complex problems. Despite its success, open questions remain as to when RL can be beneficial for industrial-scale problems. In this paper, we consider the application of MCTS with RL for optimising the reconfiguration of an MPS. We define two operational scenarios and evaluate the potential of RL in both of them. Taken more generally, our results provide context to better understand when RL can be beneficial in industrial-scale use cases.

ABSTRACT. Ensuring the trustworthiness and interpretability of machine learning models is critical to their deployment in real-world applications. Feature attribution methods have gained significant attention, which provides local explanations of model predictions by attributing importance to individual input features. This study examines the generalization of feature attributions across various deep learning architectures, such as convolutional neural networks (CNNs) and vision transformers. We aim to assess the feasibility of utilizing a feature attribution method as a future detector and examine how these features can be harmonized across multiple models employing distinct architectures but trained on the same data distribution. By exploring this harmonization, we aim to develop a more coherent and optimistic understanding of feature attributions, enhancing the consistency of local explanations across diverse deep-learning models. Our findings highlight the potential for harmonized feature attribution methods to improve interpretability and foster trust in machine learning applications, regardless of the underlying architecture.