Repository for Publications and Research Data
News from the ETH Library
Recently Added
Covering Numbers for Deep ReLU Networks with Applications to Function Approximation and Nonparametric Regression
Item type: Journal Article
Ou W.; Bölcskei H. (2026)
Covering numbers of (deep) ReLU networks have been used to characterize approximation-theoretic performance, to upper-bound prediction error in nonparametric regression, and to quantify classification capacity. These results rely on covering number upper bounds obtained via explicit constructions of coverings. Lower bounds on covering numbers do not appear to be available in the literature. The present paper fills this gap by deriving tight (up to multiplicative constants) lower and upper bounds on the metric entropy (i.e., the logarithm of the covering numbers) of fully-connected networks with bounded weights, sparse networks with bounded weights, and fully-connected networks with quantized weights. The tightness of these bounds yields a fundamental understanding of the impact of sparsity, quantization, bounded versus unbounded weights, and network output truncation. Moreover, the bounds allow one to characterize fundamental limits of neural network transformation, including network compression, and lead to sharp upper bounds on the prediction error in nonparametric regression through deep networks. In particular, we remove a (log(n))6-factor from the best known sample complexity rate for estimating Lipschitz functions via deep networks, thereby establishing optimality. Finally, we identify a systematic relation between optimal nonparametric regression and optimal approximation through deep networks, unifying numerous results in the literature and revealing underlying general principles.
Advancing biosynthetic pathway discovery through short-read-directed long-read sequencing
Item type: Review Article
Gabrielli M.; Bredel A.; Paoli L.; et al. (2026)
Time span of literature: 2020–today Metagenomic methods have rapidly advanced, enabling the identification of biosynthetic pathways directly from complex microbiome data. Short-read sequencing, while accurate and cost-effective, often generates fragmented assemblies that can lead to incomplete biosynthetic gene cluster (BGC) recovery. Although long-read sequencing offers a solution to the fragmentation problems, technical requirements and higher costs have limited its scalability. Here, we examine BGC fragmentation in short-read sequencing data across large databases of metagenome-assembled genomes (MAGs) and estimate the targeted genome contiguity required to recover ‘complete’ biosynthetic gene clusters. We argue that the increasing availability of MAGs recovered from short-read metagenomes with recent advancements in ultra-low input DNA amplification for high-fidelity PacBio sequencing—now requiring as little as nanograms of DNA—can be used sequentially to boost biosynthetic pathway discovery. We demonstrate how natural products researchers can benefit from using short-read MAG comparisons to guide targeted long-read re-sequencing efforts with low amounts of input DNA and/or limited financial resources. Our analysis provides strategic recommendations for the broader scientific community on how to best leverage the strengths of short- and long-read sequencing data to efficiently allocate resources and accelerate natural product discovery.
Choreographing the Way of Water: A Computational Framework for Aquatic Robotic Art
Item type: Conference Paper
Ramachandran A.; Golling C.; Burmester S.; et al. (2026)
Robotic choreography in open water is governed by nonlinear fluid dynamics, which impose significant challenges due to environmental disturbances and nonlinear system dynamics. This paper presents the cyber-physical architecture of Way of Water, a vertically integrated framework that orchestrates a fleet of autonomous surface vessels as a distributed choreographic platform. Moving beyond the surface-pixel paradigm, these vessels use laminar nozzles and multi-zone lighting to extend their expressive range from the 2D water plane into the 3D volumetric domain. Our primary contribution is the Way of Water Studio, a browser-based, timeline-compositing authoring paradigm that treats the fleet as a DAW-like instrument for music-responsive choreography. The Studio encapsulates Sequential Convex Programming for trajectory generation and Model Predictive Control for disturbance rejection presented through a visual timeline, broadening access to high-performance aquatic robotics for non-programmer artists. Grounding the Studio is the full cyber-physical stack: a custom holonomic chassis, a state-estimation and control stack tuned for the aquatic domain, and an LTE/MQTT fleet link with RTK-GPS time synchronization. We report on the system’s validation across two distinct deployments: an 18-vessel Swan Lake interpretation at Lake Zurich and an 8-vessel Time Space Existence 2025 Venice Biennale demonstration at Forte Marghera, establishing a foundational reference for the design and deployment of fluidic robotic swarms.
No-code computer vision model personalization for the recognition of manual assembly tasks
Item type: Conference Paper
Conrad J.; Filiberti D.; Ferchow J.; et al. (2026)
The growing complexity of manufacturing reinforces the importance of human operators in manual assembly. Their dexterity and adaptability make them indispensable, but these qualities also introduce variability in task execution. This variability poses a challenge for the use of computer vision in digital support systems to maintain efficiency and product quality. Deep learning-based vision models, while demonstrating great application potential in manufacturing, are highly sensitive to intra-class variance: differences in task execution, e.g., speed or the handling of parts and tools, can reduce recognition accuracy. This limits the robustness and industrial applicability of pretrained vision models in scenarios with frequently changing operators. One promising strategy to address this challenge is model personalization, in which pretrained models are retrained on operator-specific data to mitigate a negative effect of individual task execution styles. However, current approaches to personalize deep learning models require machine learning expertise that is scarce in industrial environments, creating a bottleneck for practical deployment. Automated machine learning (AutoML) has reduced expertise requirements in other domains, but existing frameworks do not meet all requirements for an application in manual manufacturing scenarios with high levels of variability. This paper presents a no-code tool for the personalization of deep learning–based vision models for the recognition of manual assembly tasks. The tool allows operators without coding skills to test pretrained models, record and annotate data, and automatically retrain them using a predefined pipeline. We evaluate the tool in a study with 20 operators performing three representative assembly tasks: The placement of sealings (Step 1), application of screw lock (Step 2), and fastening of screws (Step 3). Results show that personalized models outperform pretrained models for variable assembly tasks, achieving average performance changes of -0.04% for Step 1, +6.24% for Step 2, and +50.05% for Step 3 across all operators. The findings highlight the positive effect of model personalization for manual assembly tasks with high execution variance and the feasibility of a no-code tool to implement model personalization without requiring machine learning skills.
Differential rumen and hindgut microbiome and metabolome in Holstein female calves with divergent feed efficiency
Item type: Journal Article
Chen T.; Xiao J.; Li S.; et al. (2026)
BACKGROUND: Significant environmental problems have challenged animal agriculture, improving feed efficiency in animals has become a vital research direction for sustainable agriculture. Bacteria play a critical role in the feed efficiency of animals. However, our current understanding of bacteria communities in the gastrointestinal tract of high-feed efficiency animals and their metabolic mechanisms remains unclear. RESULTS: Twenty Holstein female calves were used in this multi-omics study that integrated metagenomic and metabolomic analyses of 20 Holstein female calves to investigate feed efficiency, as measured by residual feed intake (RFI). From an initial cohort of 84 calves, the 10 with the highest RFI (HRFI, low efficiency) and the 10 with the lowest RFI (LRFI, high efficiency) were selected at 84 days of age. Rumen fluid, feces, and serum samples from these calves were collected for subsequent analyses. We found that LRFI calves harbored rumen and fecal microbiomes with significantly different community structures and co-occurrence networks compared to HRFI calves. Multi-omics integration identified robust microbial and metabolite biomarkers discriminating RFI groups. These microbiomes were functionally linked to differential nutrient utilization, LRFI calves were characterized by enhanced starch and protein digestibility coupled with propionate-oriented fermentation, associated with key species like Erysipelotrichaceae_bacterium and Hungatella_sp. Conversely, HRFI calves showed higher fat digestibility and acetate production. Notably, serum glutamate was enriched in LRFI calves despite lower intake, correlating with potential microbial metabolites (ribitol, taurine). Subsequent validation confirmed that glutamate supplementation in mice improved nitrogen metabolism and gut barrier function. CONCLUSIONS: In summary, this multi-omics study reveals that high feed efficiency in calves is associated with distinct microbial ecosystems characterized by functions such as starch degradation and propionate production, where glutamate metabolism serves as a central node. Video Abstract.