Thiago S. Gouvêa

Aliki Anagnostopoulou, Thiago Gouvea, Daniel Sonntag

Large language models (LLMs) and large multimodal models (LMMs) have significantly impacted the AI community, industry, and various economic sectors. In journalism, integrating AI poses unique challenges and opportunities, particularly in enhancing the quality and efficiency of news reporting. This study explores how LLMs and LMMs can assist journalistic practice by generating contextualised captions for images accompanying news articles. We conducted experiments using the GoodNews dataset to evaluate the ability of LMMs (BLIP-2, GPT-4v, or LLaVA) to incorporate one of two types of context: entire news articles, or extracted named entities. In addition, we compared their performance to a two-stage pipeline composed of a captioning model (BLIP-2, OFA, or ViT-GPT2) with post-hoc contextualisation with LLMs (GPT-4 or LLaMA). We assess a diversity of models, and we find that while the choice of contextualisation model is a significant factor for the two-stage pipelines, this is not the case in the LMMs, where smaller, open-source models perform well compared to proprietary, GPT-powered ones. Additionally, we found that controlling the amount of provided context enhances performance. These results highlight the limitations of a fully automated approach and underscore the necessity for an interactive, human-in-the-loop strategy.

Hannes Kath, Thiago S. Gouvêa, Daniel Sonntag

Machine learning models excel with abundant annotated data, but annotation is often costly and time-intensive. Active learning (AL) aims to improve the performance-to-annotation ratio by using query methods (QMs) to iteratively select the most informative samples. While AL research focuses mainly on QM development, the evaluation of this iterative process lacks appropriate performance metrics. This work reviews eight years of AL evaluation literature and formally introduces the speed-up factor, a quantitative multi-iteration QM performance metric that indicates the fraction of samples needed to match random sampling performance. Using four datasets from diverse domains and seven QMs of various types, we empirically evaluate the speed-up factor and compare it with state-of-the-art AL performance metrics. The results confirm the assumptions underlying the speed-up factor, demonstrate its accuracy in capturing the described fraction, and reveal its superior stability across iterations.

Daniel Sonntag, Michael Barz, Thiago Gouvêa

This DFKI technical report presents the anatomy of the No-IDLE prototype system (funded by the German Federal Ministry of Education and Research) that provides not only basic and fundamental research in interactive machine learning, but also reveals deeper insights into users' behaviours, needs, and goals. Machine learning and deep learning should become accessible to millions of end users. No-IDLE's goals and scienfific challenges centre around the desire to increase the reach of interactive deep learning solutions for non-experts in machine learning. One of the key innovations described in this technical report is a methodology for interactive machine learning combined with multimodal interaction which will become central when we start interacting with semi-intelligent machines in the upcoming area of neural networks and large language models.

Hannes Kath, Bengt Lüers, Thiago S. Gouvêa et al.

Deep learning is ubiquitous, but its lack of transparency limits its impact on several potential application areas. We demonstrate a virtual reality tool for automating the process of assigning data inputs to different categories. A dataset is represented as a cloud of points in virtual space. The user explores the cloud through movement and uses hand gestures to categorise portions of the cloud. This triggers gradual movements in the cloud: points of the same category are attracted to each other, different groups are pushed apart, while points are globally distributed in a way that utilises the entire space. The space, time, and forces observed in virtual reality can be mapped to well-defined machine learning concepts, namely the latent space, the training epochs and the backpropagation. Our tool illustrates how the inner workings of deep neural networks can be made tangible and transparent. We expect this approach to accelerate the autonomous development of deep learning applications by end users in novel areas.

Hannes Kath, Thiago S. Gouvêa, Daniel Sonntag

The impact of machine learning (ML) in many fields of application is constrained by lack of annotated data. Among existing tools for ML-assisted data annotation, one little explored tool type relies on an analogy between the coordinates of a graphical user interface and the latent space of a neural network for interaction through direct manipulation. In the present work, we 1) expand the paradigm by proposing two new analogies: time and force as reflecting iterations and gradients of network training; 2) propose a network model for learning a compact graphical representation of the data that takes into account both its internal structure and user provided annotations; and 3) investigate the impact of model hyperparameters on the learned graphical representations of the data, identifying candidate model variants for a future user study.