Samuel Tovey

Konstantin Nikolaou, Jonas Scheunemann, Sven Krippendorf et al.

Neural scaling laws describe predictable power-law relationships between model size, dataset size, compute, and performance. While these laws guide the development of modern foundation models, the mechanisms underpinning them remain poorly understood, in part due to the absence of scalable analysis tools. To close this gap, we introduce "spectral position": a scalable measure of which eigenvalues of the empirical neural tangent kernel (eNTK) currently drive loss reduction. Applying this measure to scaling experiments, we find that spectral position decreases throughout training: learning shifts from dominant eigenmodes into the spectral tail. Larger models reach further into the tail than smaller models, revealing a size-dependent capacity we call "spectral reach". This suggests why larger models achieve lower losses: they sustain learning on weak spectral signals inaccessible to smaller models. We further identify feature learning as a key enabler of spectral reach. It adaptively amplifies gradient magnitudes as learning advances, sustaining progress where frozen representations stall. This points to concrete interventions through architecture and optimizer design.

Samuel Tovey, David Zimmer, Christoph Lohrmann et al.

Multi-Agent Reinforcement Learning (MARL) is a promising candidate for realizing efficient control of microscopic particles, of which micro-robots are a subset. However, the microscopic particles' environment presents unique challenges, such as Brownian motion at sufficiently small length-scales. In this work, we explore the role of temperature in the emergence and efficacy of strategies in MARL systems using particle-based Langevin molecular dynamics simulations as a realistic representation of micro-scale environments. To this end, we perform experiments on two different multi-agent tasks in microscopic environments at different temperatures, detecting the source of a concentration gradient and rotation of a rod. We find that at higher temperatures, the RL agents identify new strategies for achieving these tasks, highlighting the importance of understanding this regime and providing insight into optimal training strategies for bridging the generalization gap between simulation and reality. We also introduce a novel Python package for studying microscopic agents using reinforcement learning (RL) to accompany our results.

Samuel Tovey, Christoph Lohrmann, Tobias Merkt et al.

This work introduces SwarmRL, a Python package designed to study intelligent active particles. SwarmRL provides an easy-to-use interface for developing models to control microscopic colloids using classical control and deep reinforcement learning approaches. These models may be deployed in simulations or real-world environments under a common framework. We explain the structure of the software and its key features and demonstrate how it can be used to accelerate research. With SwarmRL, we aim to streamline research into micro-robotic control while bridging the gap between experimental and simulation-driven sciences. SwarmRL is available open-source on GitHub at https://github.com/SwarmRL/SwarmRL.

Fabian Zills, Moritz Schäfer, Samuel Tovey et al.

The past decade has seen tremendous breakthroughs in computation and there is no indication that this will slow any time soon. Machine learning, large-scale computing resources, and increased industry focus have resulted in rising investments in computer-driven solutions for data management, simulations, and model generation. However, with this growth in computation has come an even larger expansion of data and with it, complexity in data storage, sharing, and tracking. In this work, we introduce ZnTrack, a Python-driven data versioning tool. ZnTrack builds upon established version control systems to provide a user-friendly and easy-to-use interface for tracking parameters in experiments, designing workflows, and storing and sharing data. From this ability to reduce large datasets to a simple Python script emerges the concept of Data as Code, a core component of the work presented here and an undoubtedly important concept as the age of computation continues to evolve. ZnTrack offers an open-source, FAIR data compatible Python package to enable users to harness these concepts of the future.

Samuel Tovey, Christoph Lohrmann, Christian Holm

Reinforcement learning (RL) is a flexible and efficient method for programming micro-robots in complex environments. Here we investigate whether reinforcement learning can provide insights into biological systems when trained to perform chemotaxis. Namely, whether we can learn about how intelligent agents process given information in order to swim towards a target. We run simulations covering a range of agent shapes, sizes, and swim speeds to determine if the physical constraints on biological swimmers, namely Brownian motion, lead to regions where reinforcement learners' training fails. We find that the RL agents can perform chemotaxis as soon as it is physically possible and, in some cases, even before the active swimming overpowers the stochastic environment. We study the efficiency of the emergent policy and identify convergence in agent size and swim speeds. Finally, we study the strategy adopted by the reinforcement learning algorithm to explain how the agents perform their tasks. To this end, we identify three emerging dominant strategies and several rare approaches taken. These strategies, whilst producing almost identical trajectories in simulation, are distinct and give insight into the possible mechanisms behind which biological agents explore their environment and respond to changing conditions.

Samuel Tovey, Julian Hoßbach, Sandro Kuppel et al.

A device capable of performing real time classification of proteins in a clinical setting would allow for inexpensive and rapid disease diagnosis. One such candidate for this technology are nanopore devices. These devices work by measuring a current signal that arises when a protein or peptide enters a nanometer-length-scale pore. Should this current be uniquely related to the structure of the peptide and its interactions with the pore, the signals can be used to perform identification. While such a method would allow for real time identification of peptides and proteins in a clinical setting, to date, the complexities of these signals limit their accuracy. In this work, we tackle the issue of classification by converting the current signals into scaleogram images via wavelet transforms, capturing amplitude, frequency, and time information in a modality well-suited to machine learning algorithms. When tested on 42 peptides, our method achieved a classification accuracy of ~$81\,\%$, setting a new state-of-the-art in the field and taking a step toward practical peptide/protein diagnostics at the point of care. In addition, we demonstrate model transfer techniques that will be critical when deploying these models into real hardware, paving the way to a new method for real-time disease diagnosis.

Konstantin Nikolaou, Sven Krippendorf, Samuel Tovey et al.

Scaling laws offer valuable insights into the relationship between neural network performance and computational cost, yet their underlying mechanisms remain poorly understood. In this work, we empirically analyze how neural networks behave under data and model scaling through the lens of the neural tangent kernel (NTK). This analysis establishes a link between performance scaling and the internal dynamics of neural networks. Our findings of standard vision tasks show that similar performance scaling exponents can occur even though the internal model dynamics show opposite behavior. This demonstrates that performance scaling alone is insufficient for understanding the underlying mechanisms of neural networks. We also address a previously unresolved issue in neural scaling: how convergence to the infinite-width limit affects scaling behavior in finite-width models. To this end, we investigate how feature learning is lost as the model width increases and quantify the transition between kernel-driven and feature-driven scaling regimes. We identify the maximum model width that supports feature learning, which, in our setups, we find to be more than ten times smaller than typical large language model widths.

Samuel Tovey, Sven Krippendorf, Konstantin Nikolaou et al.

A theory of neural networks (NNs) built upon collective variables would provide scientists with the tools to better understand the learning process at every stage. In this work, we introduce two such variables, the entropy and the trace of the empirical neural tangent kernel (NTK) built on the training data passed to the model. We empirically analyze the NN performance in the context of these variables and find that there exists correlation between the starting entropy, the trace of the NTK, and the generalization of the model computed after training is complete. This framework is then applied to the problem of optimal data selection for the training of NNs. To this end, random network distillation (RND) is used as a means of selecting training data which is then compared with random selection of data. It is shown that not only does RND select data-sets capable of outperforming random selection, but that the collective variables associated with the RND data-sets are larger than those of the randomly selected sets. The results of this investigation provide a stable ground from which the selection of data for NN training can be driven by this phenomenological framework.