Pola Schwöbel

Pola Schwöbel, Luca Franceschi, Muhammad Bilal Zafar et al.

fmeval is an open source library to evaluate large language models (LLMs) in a range of tasks. It helps practitioners evaluate their model for task performance and along multiple responsible AI dimensions. This paper presents the library and exposes its underlying design principles: simplicity, coverage, extensibility and performance. We then present how these were implemented in the scientific and engineering choices taken when developing fmeval. A case study demonstrates a typical use case for the library: picking a suitable model for a question answering task. We close by discussing limitations and further work in the development of the library. fmeval can be found at https://github.com/aws/fmeval.

Pola Schwöbel, Jacek Golebiowski, Michele Donini et al.

Large language models (LLMs) encode vast amounts of world knowledge. However, since these models are trained on large swaths of internet data, they are at risk of inordinately capturing information about dominant groups. This imbalance can propagate into generated language. In this work, we study and operationalise a form of geographical erasure, wherein language models underpredict certain countries. We demonstrate consistent instances of erasure across a range of LLMs. We discover that erasure strongly correlates with low frequencies of country mentions in the training corpus. Lastly, we mitigate erasure by finetuning using a custom objective.

Pola Schwöbel, Peter Remmers

In recent years, the idea of formalising and modelling fairness for algorithmic decision making (ADM) has advanced to a point of sophisticated specialisation. However, the relations between technical (formalised) and ethical discourse on fairness are not always clear and productive. Arguing for an alternative perspective, we review existing fairness metrics and discuss some common issues. For instance, the fairness of procedures and distributions is often formalised and discussed statically, disregarding both structural preconditions of the status quo and downstream effects of a given intervention. We then introduce dynamic fairness modelling, a more comprehensive approach that realigns formal fairness metrics with arguments from the ethical discourse. A dynamic fairness model incorporates (1) ethical goals, (2) formal metrics to quantify decision procedures and outcomes and (3) mid-term or long-term downstream effects. By contextualising these elements of fairness-related processes, dynamic fairness modelling explicates formerly latent ethical aspects and thereby provides a helpful tool to navigate trade-offs between different fairness interventions. To illustrate the framework, we discuss an example application -- the current European efforts to increase the number of women on company boards, e.g. via quota solutions -- and present early technical work that fits within our framework.

Jan Malte Lichtenberg, Alexander Buchholz, Pola Schwöbel

The issue of popularity bias -- where popular items are disproportionately recommended, overshadowing less popular but potentially relevant items -- remains a significant challenge in recommender systems. Recent advancements have seen the integration of general-purpose Large Language Models (LLMs) into the architecture of such systems. This integration raises concerns that it might exacerbate popularity bias, given that the LLM's training data is likely dominated by popular items. However, it simultaneously presents a novel opportunity to address the bias via prompt tuning. Our study explores this dichotomy, examining whether LLMs contribute to or can alleviate popularity bias in recommender systems. We introduce a principled way to measure popularity bias by discussing existing metrics and proposing a novel metric that fulfills a series of desiderata. Based on our new metric, we compare a simple LLM-based recommender to traditional recommender systems on a movie recommendation task. We find that the LLM recommender exhibits less popularity bias, even without any explicit mitigation.

Pola Schwöbel, Martin Jørgensen, Sebastian W. Ober et al.

Data augmentation is often used to incorporate inductive biases into models. Traditionally, these are hand-crafted and tuned with cross validation. The Bayesian paradigm for model selection provides a path towards end-to-end learning of invariances using only the training data, by optimising the marginal likelihood. Computing the marginal likelihood is hard for neural networks, but success with tractable approaches that compute the marginal likelihood for the last layer only raises the question of whether this convenient approach might be employed for learning invariances. We show partial success on standard benchmarks, in the low-data regime and on a medical imaging dataset by designing a custom optimisation routine. Introducing a new lower bound to the marginal likelihood allows us to perform inference for a larger class of likelihood functions than before. On the other hand, we demonstrate failure modes on the CIFAR10 dataset, where the last layer approximation is not sufficient due to the increased complexity of our neural network. Our results indicate that once more sophisticated approximations become available the marginal likelihood is a promising approach for invariance learning in neural networks.

Pola Schwöbel, Frederik Warburg, Martin Jørgensen et al.

Spatial Transformer Networks (STNs) estimate image transformations that can improve downstream tasks by `zooming in' on relevant regions in an image. However, STNs are hard to train and sensitive to mis-predictions of transformations. To circumvent these limitations, we propose a probabilistic extension that estimates a stochastic transformation rather than a deterministic one. Marginalizing transformations allows us to consider each image at multiple poses, which makes the localization task easier and the training more robust. As an additional benefit, the stochastic transformations act as a localized, learned data augmentation that improves the downstream tasks. We show across standard imaging benchmarks and on a challenging real-world dataset that these two properties lead to improved classification performance, robustness and model calibration. We further demonstrate that the approach generalizes to non-visual domains by improving model performance on time-series data.