Anders Kirk Uhrenholt
I'm an AI Scientist with a PhD from the University of Glasgow and 8 years of experience across industry and research.
I'm currently Principal Agent Architect at Kallidin, an early-stage AI company building the Autonomous Data Office: a multi-agent system that automates the work of an enterprise data team. My focus is on the architecture of the agents themselves, and specifically on making their output reliable enough to support real business decisions.
Before that, I spent four years as an Applied Scientist at Amazon, designing deep neural recommender systems serving hundreds of millions of customers and leading the integration of generative AI into traditional ML models for improved customer protection and engagement.
My research background is in probabilistic machine learning (Gaussian processes, Bayesian optimisation, and variational inference), with publications at UAI, AISTATS, and ECIR. That work was fundamentally about quantifying uncertainty: getting a model to know how much to trust its own output. The same problem turns out to be central to multi-agent systems, where a single agent's unreliable output can propagate through an entire pipeline. Making that judgement of trust explicit and measurable, rather than assumed, is what I focus on now.
I care about building systems that are principled, reliable, and grounded in genuine understanding rather than scale alone.
Publications
Improving Minimax Group Fairness in Sequential Recommendation
The 47th European Conference on Information Retrieval, IR for Good track
Training sequential recommenders such as SASRec with uniform sample weights achieves good overall performance but can fall short on specific user groups. One such example is popularity bias, where mainstream users receive better recommendations than niche content viewers. To improve recommendation quality across diverse user groups, we explore three Distributionally Robust Optimization(DRO) methods: Group DRO, Streaming DRO, and Conditional Value at Risk (CVaR) DRO. While Group and Streaming DRO rely on group annotations and struggle with users belonging to multiple groups, CVaR does not require such annotations and can naturally handle overlapping groups. In experiments on two real-world datasets, we show that the DRO methods outperform standard training, with CVaR delivering the best results. Additionally, we find that Group and Streaming DRO are sensitive to the choice of group used for loss computation. Our contributions include (i) a novel application of CVaR to recommenders, (ii) showing that the DRO methods improve group metrics as well as overall performance, and (iii) demonstrating CVaR's effectiveness in the practical scenario of intersecting user groups.
PhD thesis: Assumptions and Efficiency in Gaussian Process Modelling
This thesis focuses on advancing resource-efficient models within probabilistic machine learning, particularly through Gaussian processes (GPs). While modern machine learning often relies on large datasets and commensurate computational power to optimize complex models, this work emphasizes the need for parsimonious approaches that carefully taylor model design to the target task in order to improve data efficiency. By examining the assumptions underlying GP models, this thesis proposes modifications that enhance both data and computational efficiency without compromising model performance. The research addresses three key areas: sparse variational inference, Bayesian optimization for target vector estimation, and preference learning with multiple users. For each, existing assumptions are revisited to improve efficiency, resulting in advancements like probabilistic inducing point selection, novel acquisition functions, and the incorporation of latent user features in preference learning. These findings offer new pathways for designing machine learning models that balance complexity with resource constraints, making them more applicable to a wide range of practical scenarios.
Probabilistic Selection of Inducing Points in Sparse Gaussian Processes
The 37th Conference on Uncertainty in Articifial Intelligence.
Sparse Gaussian processes and various extensions thereof are enabled through inducing points, that simultaneously bottleneck the predictive capacity and act as the main contributor towards model complexity. However, the number of inducing points is generally not associated with uncertainty which prevents us from applying the apparatus of Bayesian reasoning in identifying an appropriate tradeoff. In this work we place a point process prior on the inducing points and approximate the associated posterior through stochastic variational inference. By letting the prior encourage a moderate number of inducing points, we enable the model to learn which and how many points to utilise. We experimentally show that fewer inducing points are preferred by the model as the points become less informative, and further demonstrate how the method can be applied in deep Gaussian processes and latent variable modelling.
Odd-One-Out Representation Learning
Object Representations for Learning and Reasoning (NeurIPS workshop)
The effective application of representation learning to real-world
problems requires both techniques for learning useful representations,
and also robust ways to evaluate properties of representations. Recent
work in disentangled representation learning has shown that
unsupervised representation learning approaches rely on fully
supervised disentanglement metrics, which assume access to labels for
ground-truth factors of variation. In many real-world cases
ground-truth factors are expensive to collect, or difficult to model,
such as for perception. Here we empirically show that a
weakly-supervised downstream task based on odd-one-out observations is
suitable for model selection by observing high correlation on a
difficult downstream abstract visual reasoning task. We also show that
a bespoke metric-learning VAE model which performs highly on this task
also out-performs other standard unsupervised and a weakly-supervised
disentanglement model across several metrics.
Efficient Bayesian Optimization for Target Vector Estimation
The 22nd International Conference on Artificial Intelligence and Statistics.
We consider the problem of estimating a target vector by querying an unknown multi-output function which is stochastic and expensive to evaluate. Through sequential experimental design the aim is to minimize the squared Euclidean distance between the output of the function and the target vector. Applying standard single-objective Bayesian optimization to this problem is both wasteful, since individual output components are never observed, and imprecise since the predictive distribution for new inputs will be symmetric and have negative support. We address this issue by proposing a Gaussian process model that considers the individual function outputs and derive a distribution over the resulting 2-norm. Furthermore we derive computationally efficient acquisition functions and evaluate the resulting optimization framework on several synthetic problems and a real-world problem. The results demonstrate a significant improvement over Bayesian optimization based on both standard and warped Gaussian processes.
Exploring Visualisation of Channel Activity, Levels and EQ For User Interfaces Implementing the Stage Metaphor for Music Mixing
Proceedings of the 2nd AES Workshop on Intelligent Music Production. Vol. 13.
This short précis outlines a collection of different strategies for visualising simple audio features for a GUIbased audio mixing interface that uses the stage metaphor control scheme. Audio features such as activity, loudness and spectral centroid are extracted in real-time and mapped to different visual cues that can be adapted to the circular widgets most often found in implementations of the stage metaphor. An initial evaluation suggests that while the visualisations are generally intuitive and provide information about activity of audio channels, they are not used directly. When implementing these kinds of dynamic graphical visualisations it is thus important to consider how intrusive they are compared to their usefulness in a real mixing context.
Towards Improving Overview and Metering through Visualization and Dynamic Query Filters for User Interfaces Implementing the Stage Metaphor for Music Mixing
Audio Engineering Society Convention 141. Audio Engineering Society.
This paper deals with challenges involved with implementing the stage metaphor control scheme for mixing music. Recent studies suggest that the stage metaphor outperforms the traditional channel-strip metaphor in several different ways. However, the implementation of the stage metaphor poses issues including clutter, lack of overview and monitoring of levels, and EQ. Drawing upon suggestions in recent studies, the paper describes the implementation of a stage metaphor prototype incorporating several features for dealing with these issues, including level and EQ monitoring using brightness, shape, and size. Moreover we explore the potential of using Dynamic Query filtering for localizing channels with certain properties of interest. Finally, an explorative user evaluation compares different variations of the prototype, leading to a discussion of the importance of each feature.