Damian Borth

Konstantin Schürholt, Boris Knyazev, Xavier Giró-i-Nieto et al. · berkeley

Learning representations of neural network weights given a model zoo is an emerging and challenging area with many potential applications from model inspection, to neural architecture search or knowledge distillation. Recently, an autoencoder trained on a model zoo was able to learn a hyper-representation, which captures intrinsic and extrinsic properties of the models in the zoo. In this work, we extend hyper-representations for generative use to sample new model weights. We propose layer-wise loss normalization which we demonstrate is key to generate high-performing models and several sampling methods based on the topology of hyper-representations. The models generated using our methods are diverse, performant and capable to outperform strong baselines as evaluated on several downstream tasks: initialization, ensemble sampling and transfer learning. Our results indicate the potential of knowledge aggregation from model zoos to new models via hyper-representations thereby paving the avenue for novel research directions.

Konstantin Schürholt, Diyar Taskiran, Boris Knyazev et al. · berkeley

In the last years, neural networks (NN) have evolved from laboratory environments to the state-of-the-art for many real-world problems. It was shown that NN models (i.e., their weights and biases) evolve on unique trajectories in weight space during training. Following, a population of such neural network models (referred to as model zoo) would form structures in weight space. We think that the geometry, curvature and smoothness of these structures contain information about the state of training and can reveal latent properties of individual models. With such model zoos, one could investigate novel approaches for (i) model analysis, (ii) discover unknown learning dynamics, (iii) learn rich representations of such populations, or (iv) exploit the model zoos for generative modelling of NN weights and biases. Unfortunately, the lack of standardized model zoos and available benchmarks significantly increases the friction for further research about populations of NNs. With this work, we publish a novel dataset of model zoos containing systematically generated and diverse populations of NN models for further research. In total the proposed model zoo dataset is based on eight image datasets, consists of 27 model zoos trained with varying hyperparameter combinations and includes 50'360 unique NN models as well as their sparsified twins, resulting in over 3'844'360 collected model states. Additionally, to the model zoo data we provide an in-depth analysis of the zoos and provide benchmarks for multiple downstream tasks. The dataset can be found at www.modelzoos.cc.

Marco Schreyer, Hamed Hemati, Damian Borth et al. · berkeley

The International Standards on Auditing require auditors to collect reasonable assurance that financial statements are free of material misstatement. At the same time, a central objective of Continuous Assurance is the real-time assessment of digital accounting journal entries. Recently, driven by the advances in artificial intelligence, Deep Learning techniques have emerged in financial auditing to examine vast quantities of accounting data. However, learning highly adaptive audit models in decentralised and dynamic settings remains challenging. It requires the study of data distribution shifts over multiple clients and time periods. In this work, we propose a Federated Continual Learning framework enabling auditors to learn audit models from decentral clients continuously. We evaluate the framework's ability to detect accounting anomalies in common scenarios of organizational activity. Our empirical results, using real-world datasets and combined federated continual learning strategies, demonstrate the learned model's ability to detect anomalies in audit settings of data distribution shifts.

Marco Schreyer, Timur Sattarov, Damian Borth · berkeley

The ongoing 'digital transformation' fundamentally changes audit evidence's nature, recording, and volume. Nowadays, the International Standards on Auditing (ISA) requires auditors to examine vast volumes of a financial statement's underlying digital accounting records. As a result, audit firms also 'digitize' their analytical capabilities and invest in Deep Learning (DL), a successful sub-discipline of Machine Learning. The application of DL offers the ability to learn specialized audit models from data of multiple clients, e.g., organizations operating in the same industry or jurisdiction. In general, regulations require auditors to adhere to strict data confidentiality measures. At the same time, recent intriguing discoveries showed that large-scale DL models are vulnerable to leaking sensitive training data information. Today, it often remains unclear how audit firms can apply DL models while complying with data protection regulations. In this work, we propose a Federated Learning framework to train DL models on auditing relevant accounting data of multiple clients. The framework encompasses Differential Privacy and Split Learning capabilities to mitigate data confidentiality risks at model inference. We evaluate our approach to detect accounting anomalies in three real-world datasets of city payments. Our results provide empirical evidence that auditors can benefit from DL models that accumulate knowledge from multiple sources of proprietary client data.

Michael Mommert, Nicolas Kesseli, Joëlle Hanna et al. · berkeley

Deep learning methods have proven to be a powerful tool in the analysis of large amounts of complex Earth observation data. However, while Earth observation data are multi-modal in most cases, only single or few modalities are typically considered. In this work, we present the ben-ge dataset, which supplements the BigEarthNet-MM dataset by compiling freely and globally available geographical and environmental data. Based on this dataset, we showcase the value of combining different data modalities for the downstream tasks of patch-based land-use/land-cover classification and land-use/land-cover segmentation. ben-ge is freely available and expected to serve as a test bed for fully supervised and self-supervised Earth observation applications.

Shijun Wang, Jón Guðnason, Damian Borth · berkeley

State-of-the-art Text-To-Speech (TTS) models are capable of producing high-quality speech. The generated speech, however, is usually neutral in emotional expression, whereas very often one would want fine-grained emotional control of words or phonemes. Although still challenging, the first TTS models have been recently proposed that are able to control voice by manually assigning emotion intensity. Unfortunately, due to the neglect of intra-class distance, the intensity differences are often unrecognizable. In this paper, we propose a fine-grained controllable emotional TTS, that considers both inter- and intra-class distances and be able to synthesize speech with recognizable intensity difference. Our subjective and objective experiments demonstrate that our model exceeds two state-of-the-art controllable TTS models for controllability, emotion expressiveness and naturalness.

Hamed Hemati, Andrea Cossu, Antonio Carta et al. · berkeley

Real-world data streams naturally include the repetition of previous concepts. From a Continual Learning (CL) perspective, repetition is a property of the environment and, unlike replay, cannot be controlled by the agent. Nowadays, the Class-Incremental (CI) scenario represents the leading test-bed for assessing and comparing CL strategies. This scenario type is very easy to use, but it never allows revisiting previously seen classes, thus completely neglecting the role of repetition. We focus on the family of Class-Incremental with Repetition (CIR) scenario, where repetition is embedded in the definition of the stream. We propose two stochastic stream generators that produce a wide range of CIR streams starting from a single dataset and a few interpretable control parameters. We conduct the first comprehensive evaluation of repetition in CL by studying the behavior of existing CL strategies under different CIR streams. We then present a novel replay strategy that exploits repetition and counteracts the natural imbalance present in the stream. On both CIFAR100 and TinyImageNet, our strategy outperforms other replay approaches, which are not designed for environments with repetition.

Konstantin Schürholt, Boris Knyazev, Xavier Giró-i-Nieto et al. · berkeley

Learning representations of neural network weights given a model zoo is an emerging and challenging area with many potential applications from model inspection, to neural architecture search or knowledge distillation. Recently, an autoencoder trained on a model zoo was able to learn a hyper-representation, which captures intrinsic and extrinsic properties of the models in the zoo. In this work, we extend hyper-representations for generative use to sample new model weights as pre-training. We propose layer-wise loss normalization which we demonstrate is key to generate high-performing models and a sampling method based on the empirical density of hyper-representations. The models generated using our methods are diverse, performant and capable to outperform conventional baselines for transfer learning. Our results indicate the potential of knowledge aggregation from model zoos to new models via hyper-representations thereby paving the avenue for novel research directions.

Ricardo Müller, Marco Schreyer, Timur Sattarov et al. · berkeley

Detecting accounting anomalies is a recurrent challenge in financial statement audits. Recently, novel methods derived from Deep-Learning (DL) have been proposed to audit the large volumes of a statement's underlying accounting records. However, due to their vast number of parameters, such models exhibit the drawback of being inherently opaque. At the same time, the concealing of a model's inner workings often hinders its real-world application. This observation holds particularly true in financial audits since auditors must reasonably explain and justify their audit decisions. Nowadays, various Explainable AI (XAI) techniques have been proposed to address this challenge, e.g., SHapley Additive exPlanations (SHAP). However, in unsupervised DL as often applied in financial audits, these methods explain the model output at the level of encoded variables. As a result, the explanations of Autoencoder Neural Networks (AENNs) are often hard to comprehend by human auditors. To mitigate this drawback, we propose (RESHAPE), which explains the model output on an aggregated attribute-level. In addition, we introduce an evaluation framework to compare the versatility of XAI methods in auditing. Our experimental results show empirical evidence that RESHAPE results in versatile explanations compared to state-of-the-art baselines. We envision such attribute-level explanations as a necessary next step in the adoption of unsupervised DL techniques in financial auditing.

Hamed Hemati, Vincenzo Lomonaco, Davide Bacciu et al. · berkeley

Hypernetworks mitigate forgetting in continual learning (CL) by generating task-dependent weights and penalizing weight changes at a meta-model level. Unfortunately, generating all weights is not only computationally expensive for larger architectures, but also, it is not well understood whether generating all model weights is necessary. Inspired by latent replay methods in CL, we propose partial weight generation for the final layers of a model using hypernetworks while freezing the initial layers. With this objective, we first answer the question of how many layers can be frozen without compromising the final performance. Through several experiments, we empirically show that the number of layers that can be frozen is proportional to the distributional similarity in the CL stream. Then, to demonstrate the effectiveness of hypernetworks, we show that noisy streams can significantly impact the performance of latent replay methods, leading to increased forgetting when features from noisy experiences are replayed with old samples. In contrast, partial hypernetworks are more robust to noise by maintaining accuracy on previous experiences. Finally, we conduct experiments on the split CIFAR-100 and TinyImagenet benchmarks and compare different versions of partial hypernetworks to latent replay methods. We conclude that partial weight generation using hypernetworks is a promising solution to the problem of forgetting in neural networks. It can provide an effective balance between computation and final test accuracy in CL streams.

Shijun Wang, Jón Guðnason, Damian Borth · berkeley

Effective speech emotional representations play a key role in Speech Emotion Recognition (SER) and Emotional Text-To-Speech (TTS) tasks. However, emotional speech samples are more difficult and expensive to acquire compared with Neutral style speech, which causes one issue that most related works unfortunately neglect: imbalanced datasets. Models might overfit to the majority Neutral class and fail to produce robust and effective emotional representations. In this paper, we propose an Emotion Extractor to address this issue. We use augmentation approaches to train the model and enable it to extract effective and generalizable emotional representations from imbalanced datasets. Our empirical results show that (1) for the SER task, the proposed Emotion Extractor surpasses the state-of-the-art baseline on three imbalanced datasets; (2) the produced representations from our Emotion Extractor benefit the TTS model, and enable it to synthesize more expressive speech.

Dominik Honegger, Konstantin Schürholt, Damian Borth · berkeley

With growing size of Neural Networks (NNs), model sparsification to reduce the computational cost and memory demand for model inference has become of vital interest for both research and production. While many sparsification methods have been proposed and successfully applied on individual models, to the best of our knowledge their behavior and robustness has not yet been studied on large populations of models. With this paper, we address that gap by applying two popular sparsification methods on populations of models (so called model zoos) to create sparsified versions of the original zoos. We investigate the performance of these two methods for each zoo, compare sparsification layer-wise, and analyse agreement between original and sparsified populations. We find both methods to be very robust with magnitude pruning able outperform variational dropout with the exception of high sparsification ratios above 80%. Further, we find sparsified models agree to a high degree with their original non-sparsified counterpart, and that the performance of original and sparsified model is highly correlated. Finally, all models of the model zoos and their sparsified model twins are publicly available: modelzoos.cc.

Xiaolong Han, Zehong Wang, Bo Zhao et al.

Neural network weights are typically viewed as the end product of training, while most deep learning research focuses on data, features, and architectures. However, recent advances show that the set of all possible weight values (weight space) itself contains rich structure: pretrained models form organized distributions, exhibit symmetries, and can be embedded, compared, or even generated. Understanding such structures has tremendous impact on how neural networks are analyzed and compared, and on how knowledge is transferred across models, beyond individual training instances. This emerging research direction, which we refer to as Weight Space Learning (WSL), treats neural weights as a meaningful domain for analysis and modeling. This survey provides the first unified taxonomy of WSL. We categorize existing methods into three core dimensions: Weight Space Understanding (WSU), which studies the geometry and symmetries of weights; Weight Space Representation (WSR), which learns embeddings over model weights; and Weight Space Generation (WSG), which synthesizes new weights through hypernetworks or generative models. We further show how these developments enable practical applications, including model retrieval, continual and federated learning, neural architecture search, and data-free reconstruction. By consolidating fragmented progress under a coherent framework, this survey highlights weight space as a learnable, structured domain with growing impact across model analysis, transferring, and weight generation. We release an accompanying resource at https://github.com/Zehong-Wang/Awesome-Weight-Space-Learning.

Joelle Hanna, Damian Falk, Stella X. Yu et al.

Recent advances in remote sensing have led to an increase in the number of available foundation models; each trained on different modalities, datasets, and objectives, yet capturing only part of the vast geospatial knowledge landscape. While these models show strong results within their respective domains, their capabilities remain complementary rather than unified. Therefore, instead of choosing one model over another, we aim to combine their strengths into a single shared representation. We introduce GeoSANE, a geospatial model foundry that learns a unified neural representation from the weights of existing foundation models and task-specific models, able to generate novel neural networks weights on-demand. Given a target architecture, GeoSANE generates weights ready for finetuning for classification, segmentation, and detection tasks across multiple modalities. Models generated by GeoSANE consistently outperform their counterparts trained from scratch, match or surpass state-of-the-art remote sensing foundation models, and outperform models obtained through pruning or knowledge distillation when generating lightweight networks. Evaluations across ten diverse datasets and on GEO-Bench confirm its strong generalization capabilities. By shifting from pre-training to weight generation, GeoSANE introduces a new framework for unifying and transferring geospatial knowledge across models and tasks. Code is available at \href{https://hsg-aiml.github.io/GeoSANE/}{hsg-aiml.github.io/GeoSANE/}.

Kristófer Reynisson, Marco Schreyer, Damian Borth

Credit card fraud has significant implications at both an individual and societal level, making effective prevention essential. Current methods rely heavily on feature engineering and labeled information, both of which have significant limitations. In this work, we present GraphGuard, a novel contrastive self-supervised graph-based framework for detecting fraudulent credit card transactions. We conduct experiments on a real-world dataset and a synthetic dataset. Our results provide a promising initial direction for exploring the effectiveness of graph-based self-supervised approaches for credit card fraud detection.

Zhitong Xiong, Yi Wang, Fahong Zhang et al.

Earth observation (EO) in open-world settings presents a unique challenge: different applications rely on diverse sensor modalities, each with varying ground sampling distances, spectral ranges, and numbers of spectral bands. However, existing EO foundation models are typically tailored to specific sensor types, making them inflexible when generalizing across the heterogeneous landscape of EO data. To address this, we propose the Dynamic One-For-All (DOFA) model, a unified, multimodal foundation framework designed for diverse vision tasks in EO. Inspired by neural plasticity, DOFA utilizes a wavelength-conditioned dynamic hypernetwork to process inputs from five distinct satellite sensors flexibly. By continually pretraining on five EO modalities, DOFA achieves state-of-the-art performance across multiple downstream tasks and generalizes well to unseen modalities. Enhanced with hybrid continual pretraining, DOFA+ requires significantly fewer computational resources while outperforming counterparts trained with extensive GPU budgets. Experiments on diverse datasets highlight DOFA's potential as a foundation for general-purpose vision models in the sensor-diverse EO domain. The code and pre-trained weights are publicly available at https://github.com/zhu-xlab/DOFA.

Damian Falk, Léo Meynent, Florence Pfammatter et al.

The availability of large, structured populations of neural networks - called 'model zoos' - has led to the development of a multitude of downstream tasks ranging from model analysis, to representation learning on model weights or generative modeling of neural network parameters. However, existing model zoos are limited in size and architecture and neglect the transformer, which is among the currently most successful neural network architectures. We address this gap by introducing the first model zoo of vision transformers (ViT). To better represent recent training approaches, we develop a new blueprint for model zoo generation that encompasses both pre-training and fine-tuning steps, and publish 250 unique models. They are carefully generated with a large span of generating factors, and their diversity is validated using a thorough choice of weight-space and behavioral metrics. To further motivate the utility of our proposed dataset, we suggest multiple possible applications grounded in both extensive exploratory experiments and a number of examples from the existing literature. By extending previous lines of similar work, our model zoo allows researchers to push their model population-based methods from the small model regime to state-of-the-art architectures. We make our model zoo available at github.com/ModelZoos/ViTModelZoo.

Damian Falk, Konstantin Schürholt, Damian Borth

Re-using trained neural network models is a common strategy to reduce training cost and transfer knowledge. Weight space learning - using the weights of trained models as data modality - is a promising new field to re-use populations of pre-trained models for future tasks. Approaches in this field have demonstrated high performance both on model analysis and weight generation tasks. However, until now their learning setup requires homogeneous model zoos where all models share the same exact architecture, limiting their capability to generalize beyond the population of models they saw during training. In this work, we remove this constraint and propose a modification to a common weight space learning method to accommodate training on heterogeneous populations of models. We further investigate the resulting impact of model diversity on generating unseen neural network model weights for zero-shot knowledge transfer. Our extensive experimental evaluation shows that including models with varying underlying image datasets has a high impact on performance and generalization, for both in- and out-of-distribution settings. Code is available on github.com/HSG-AIML/MultiZoo-SANE.

Alexander Arimond, Mauro Molteni, Dominik Jany et al.

We propose the application of Transformer-based language models for classifying entity legal forms from raw legal entity names. Specifically, we employ various BERT variants and compare their performance against multiple traditional baselines. Our evaluation encompasses a substantial subset of freely available Legal Entity Identifier (LEI) data, comprising over 1.1 million legal entities from 30 different legal jurisdictions. The ground truth labels for classification per jurisdiction are taken from the Entity Legal Form (ELF) code standard (ISO 20275). Our findings demonstrate that pre-trained BERT variants outperform traditional text classification approaches in terms of F1 score, while also performing comparably well in the Macro F1 Score. Moreover, the validity of our proposal is supported by the outcome of third-party expert reviews conducted in ten selected jurisdictions. This study highlights the significant potential of Transformer-based models in advancing data standardization and data integration. The presented approaches can greatly benefit financial institutions, corporations, governments and other organizations in assessing business relationships, understanding risk exposure, and promoting effective governance.

Joelle Hanna, Linus Scheibenreif, Damian Borth

Remote sensing data is commonly used for tasks such as flood mapping, wildfire detection, or land-use studies. For each task, scientists carefully choose appropriate modalities or leverage data from purpose-built instruments. Recent work on remote sensing foundation models pre-trains computer vision models on large amounts of remote sensing data. These large-scale models tend to focus on specific modalities, often optical RGB or multispectral data. For many important applications, this introduces a mismatch between the application modalities and the pre-training data. Moreover, the large size of foundation models makes them expensive and difficult to fine-tune on typically small datasets for each task. We address this mismatch with MAPEX, a remote sensing foundation model based on mixture-of-modality experts. MAPEX is pre-trained on multi-modal remote sensing data with a novel modality-conditioned token routing mechanism that elicits modality-specific experts. To apply the model on a specific task, we propose a modality aware pruning technique, which only retains experts specialized for the task modalities. This yields efficient modality-specific models while simplifying fine-tuning and deployment for the modalities of interest. We experimentally validate MAPEX on diverse remote sensing datasets and show strong performance compared to fully supervised training and state-of-the-art remote sensing foundation models. Code is available at https://github.com/HSG-AIML/MAPEX.

Mario Michael Krell, Julia Bernd, Yifan Li et al.

Social multimedia users are increasingly sharing all kinds of data about the world. They do this for their own reasons, not to provide data for field studies-but the trend presents a great opportunity for scientists. The Yahoo Flickr Creative Commons 100 Million (YFCC100M) dataset comprises 99 million images and nearly 800 thousand videos from Flickr, all shared under Creative Commons licenses. To enable scientists to leverage these media records for field studies, we propose a new framework that extracts targeted subcorpora from the YFCC100M, in a format usable by researchers who are not experts in big data retrieval and processing. This paper discusses a number of examples from the literature-as well as some entirely new ideas-of natural and social science field studies that could be piloted, supplemented, replicated, or conducted using YFCC100M data. These examples illustrate the need for a general new open-source framework for Multimedia Big Data Field Studies. There is currently a gap between the separate aspects of what multimedia researchers have shown to be possible with consumer-produced big data and the follow-through of creating a comprehensive field study framework that supports scientists across other disciplines. To bridge this gap, we must meet several challenges. For example, the framework must handle unlabeled and noisily labeled data to produce a filtered dataset for a scientist-who naturally wants it to be both as large and as clean as possible. This requires an iterative approach that provides access to statistical summaries and refines the search by constructing new classifiers. The first phase of our framework is available as Multimedia Commons Search, an intuitive interface that enables complex search queries at a large scale...

Patrick Helber, Benjamin Bischke, Andreas Dengel et al.

In this paper, we address the challenge of land use and land cover classification using Sentinel-2 satellite images. The Sentinel-2 satellite images are openly and freely accessible provided in the Earth observation program Copernicus. We present a novel dataset based on Sentinel-2 satellite images covering 13 spectral bands and consisting out of 10 classes with in total 27,000 labeled and geo-referenced images. We provide benchmarks for this novel dataset with its spectral bands using state-of-the-art deep Convolutional Neural Network (CNNs). With the proposed novel dataset, we achieved an overall classification accuracy of 98.57%. The resulting classification system opens a gate towards a number of Earth observation applications. We demonstrate how this classification system can be used for detecting land use and land cover changes and how it can assist in improving geographical maps. The geo-referenced dataset EuroSAT is made publicly available at https://github.com/phelber/eurosat.

Timur Sattarov, Marco Schreyer, Damian Borth

Realistic synthetic tabular data generation encounters significant challenges in preserving privacy, especially when dealing with sensitive information in domains like finance and healthcare. In this paper, we introduce \textit{Federated Tabular Diffusion} (FedTabDiff) for generating high-fidelity mixed-type tabular data without centralized access to the original tabular datasets. Leveraging the strengths of \textit{Denoising Diffusion Probabilistic Models} (DDPMs), our approach addresses the inherent complexities in tabular data, such as mixed attribute types and implicit relationships. More critically, FedTabDiff realizes a decentralized learning scheme that permits multiple entities to collaboratively train a generative model while respecting data privacy and locality. We extend DDPMs into the federated setting for tabular data generation, which includes a synchronous update scheme and weighted averaging for effective model aggregation. Experimental evaluations on real-world financial and medical datasets attest to the framework's capability to produce synthetic data that maintains high fidelity, utility, privacy, and coverage.

Léo Meynent, Ivan Melev, Konstantin Schürholt et al.

The weights of neural networks (NNs) have recently gained prominence as a new data modality in machine learning, with applications ranging from accuracy and hyperparameter prediction to representation learning or weight generation. One approach to leverage NN weights involves training autoencoders (AEs), using contrastive and reconstruction losses. This allows such models to be applied to a wide variety of downstream tasks, and they demonstrate strong predictive performance and low reconstruction error. However, despite the low reconstruction error, these AEs reconstruct NN models with deteriorated performance compared to the original ones, limiting their usability with regard to model weight generation. In this paper, we identify a limitation of weight-space AEs, specifically highlighting that a structural loss, that uses the Euclidean distance between original and reconstructed weights, fails to capture some features critical for reconstructing high-performing models. We analyze the addition of a behavioral loss for training AEs in weight space, where we compare the output of the reconstructed model with that of the original one, given some common input. We show a strong synergy between structural and behavioral signals, leading to increased performance in all downstream tasks evaluated, in particular NN weights reconstruction and generation.

Zhiyuan Liang, Dongwen Tang, Yuhao Zhou et al.

Modern Parameter-Efficient Fine-Tuning (PEFT) methods such as low-rank adaptation (LoRA) reduce the cost of customizing large language models (LLMs), yet still require a separate optimization run for every downstream dataset. We introduce \textbf{Drag-and-Drop LLMs (\textit{DnD})}, a prompt-conditioned parameter generator that eliminates per-task training by mapping a handful of unlabeled task prompts directly to LoRA weight updates. A lightweight text encoder distills each prompt batch into condition embeddings, which are then transformed by a cascaded hyper-convolutional decoder into the full set of LoRA matrices. Once trained in a diverse collection of prompt-checkpoint pairs, DnD produces task-specific parameters in seconds, yielding i) up to \textbf{12,000$\times$} lower overhead than full fine-tuning, ii) average gains up to \textbf{30\%} in performance over the strongest training LoRAs on unseen common-sense reasoning, math, coding, and multimodal benchmarks, and iii) robust cross-domain generalization despite never seeing the target data or labels. Our results demonstrate that prompt-conditioned parameter generation is a viable alternative to gradient-based adaptation for rapidly specializing LLMs. Our project is available at \href{https://jerryliang24.github.io/DnD}{https://jerryliang24.github.io/DnD}.

Konstantin Schürholt, Léo Meynent, Yefan Zhou et al.

Using the weights of trained Neural Network (NN) models as data modality has recently gained traction as a research field - dubbed Weight Space Learning (WSL). Multiple recent works propose WSL methods to analyze models, evaluate methods, or synthesize weights. Weight space learning methods require populations of trained models as datasets for development and evaluation. However, existing collections of models - called `model zoos' - are unstructured or follow a rudimentary definition of diversity. In parallel, work rooted in statistical physics has identified phases and phase transitions in NN models. Models are homogeneous within the same phase but qualitatively differ from one phase to another. We combine the idea of `model zoos' with phase information to create a controlled notion of diversity in populations. We introduce 12 large-scale zoos that systematically cover known phases and vary over model architecture, size, and datasets. These datasets cover different modalities, such as computer vision, natural language processing, and scientific ML. For every model, we compute loss landscape metrics and validate full coverage of the phases. With this dataset, we provide the community with a resource with a wide range of potential applications for WSL and beyond. Evidence suggests the loss landscape phase plays a role in applications such as model training, analysis, or sparsification. We demonstrate this in an exploratory study of the downstream methods like transfer learning or model weights averaging.

Timur Sattarov, Marco Schreyer, Damian Borth

The increasing demand for privacy-preserving data analytics in various domains necessitates solutions for synthetic data generation that rigorously uphold privacy standards. We introduce the DP-FedTabDiff framework, a novel integration of Differential Privacy, Federated Learning and Denoising Diffusion Probabilistic Models designed to generate high-fidelity synthetic tabular data. This framework ensures compliance with privacy regulations while maintaining data utility. We demonstrate the effectiveness of DP-FedTabDiff on multiple real-world mixed-type tabular datasets, achieving significant improvements in privacy guarantees without compromising data quality. Our empirical evaluations reveal the optimal trade-offs between privacy budgets, client configurations, and federated optimization strategies. The results affirm the potential of DP-FedTabDiff to enable secure data sharing and analytics in highly regulated domains, paving the way for further advances in federated learning and privacy-preserving data synthesis.

Ruben Gonzalez Avilés, Linus Scheibenreif, Nassim Ait Ali Braham et al.

Hyperspectral imaging provides detailed spectral information and holds significant potential for monitoring of greenhouse gases (GHGs). However, its application is constrained by limited spatial coverage and infrequent revisit times. In contrast, multispectral imaging offers broader spatial and temporal coverage but often lacks the spectral detail that can enhance GHG detection. To address these challenges, this study proposes a spectral transformer model that synthesizes hyperspectral data from multispectral inputs. The model is pre-trained via a band-wise masked autoencoder and subsequently fine-tuned on spatio-temporally aligned multispectral-hyperspectral image pairs. The resulting synthetic hyperspectral data retain the spatial and temporal benefits of multispectral imagery and improve GHG prediction accuracy relative to using multispectral data alone. This approach effectively bridges the trade-off between spectral resolution and coverage, highlighting its potential to advance atmospheric monitoring by combining the strengths of hyperspectral and multispectral systems with self-supervised deep learning.

Damian Falk, Konstantin Schürholt, Konstantinos Tzevelekakis et al.

The weights of neural networks have emerged as a novel data modality, giving rise to the field of weight space learning. A central challenge in this area is that learning meaningful representations of weights typically requires large, carefully constructed collections of trained models, typically referred to as model zoos. These model zoos are often trained ad-hoc, requiring large computational resources, constraining the learned weight space representations in scale and flexibility. In this work, we drop this requirement by training a weight space learning backbone on arbitrary models downloaded from large, unstructured model repositories such as Hugging Face. Unlike curated model zoos, these repositories contain highly heterogeneous models: they vary in architecture and dataset, and are largely undocumented. To address the methodological challenges posed by such heterogeneity, we propose a new weight space backbone designed to handle unstructured model populations. We demonstrate that weight space representations trained on models from Hugging Face achieve strong performance, often outperforming backbones trained on laboratory-generated model zoos. Finally, we show that the diversity of the model weights in our training set allows our weight space model to generalize to unseen data modalities. By demonstrating that high-quality weight space representations can be learned in the wild, we show that curated model zoos are not indispensable, thereby overcoming a strong limitation currently faced by the weight space learning community.

Timur Sattarov, Marco Schreyer, Damian Borth

Anomaly detection in tabular data remains challenging due to complex feature interactions and the scarcity of anomalous examples. Denoising autoencoders rely on fixed-magnitude noise, limiting adaptability to diverse data distributions. Diffusion models introduce scheduled noise and iterative denoising, but lack explicit reconstruction mappings. We propose the Diffusion-Scheduled Denoising Autoencoder (DDAE), a framework that integrates diffusion-based noise scheduling and contrastive learning into the encoding process to improve anomaly detection. We evaluated DDAE on 57 datasets from ADBench. Our method outperforms in semi-supervised settings and achieves competitive results in unsupervised settings, improving PR-AUC by up to 65% (9%) and ROC-AUC by 16% (6%) over state-of-the-art autoencoder (diffusion) model baselines. We observed that higher noise levels benefit unsupervised training, while lower noise with linear scheduling is optimal in semi-supervised settings. These findings underscore the importance of principled noise strategies in tabular anomaly detection.

Joelle Hanna, Damian Borth

Weakly Supervised Semantic Segmentation (WSSS) is a challenging problem that has been extensively studied in recent years. Traditional approaches often rely on external modules like Class Activation Maps to highlight regions of interest and generate pseudo segmentation masks. In this work, we propose an end-to-end method that directly utilizes the attention maps learned by a Vision Transformer (ViT) for WSSS. We propose training a sparse ViT with multiple [CLS] tokens (one for each class), using a random masking strategy to promote [CLS] token - class assignment. At inference time, we aggregate the different self-attention maps of each [CLS] token corresponding to the predicted labels to generate pseudo segmentation masks. Our proposed approach enhances the interpretability of self-attention maps and ensures accurate class assignments. Extensive experiments on two standard benchmarks and three specialized datasets demonstrate that our method generates accurate pseudo-masks, outperforming related works. Those pseudo-masks can be used to train a segmentation model which achieves results comparable to fully-supervised models, significantly reducing the need for fine-grained labeled data.

Ruben Gonzalez Avilés, Linus Scheibenreif, Damian Borth

This paper addresses the critical environmental challenge of estimating ambient Nitrogen Dioxide (NO$_2$) concentrations, a key issue in public health and environmental policy. Existing methods for satellite-based air pollution estimation model the relationship between satellite and in-situ measurements at select point locations. While these approaches have advanced our ability to provide air quality estimations on a global scale, they come with inherent limitations. The most notable limitation is the computational intensity required for generating comprehensive estimates over extensive areas. Motivated by these limitations, this study introduces a novel dense estimation technique. Our approach seeks to balance the accuracy of high-resolution estimates with the practicality of computational constraints, thereby enabling efficient and scalable global environmental assessment. By utilizing a uniformly random offset sampling strategy, our method disperses the ground truth data pixel location evenly across a larger patch. At inference, the dense estimation method can then generate a grid of estimates in a single step, significantly reducing the computational resources required to provide estimates for larger areas. Notably, our approach also surpasses the results of existing point-wise methods by a significant margin of $9.45\%$, achieving a Mean Absolute Error (MAE) of $4.98\ μ\text{g}/\text{m}^3$. This demonstrates both high accuracy and computational efficiency, highlighting the applicability of our method for global environmental assessment. Furthermore, we showcase the method's adaptability and robustness by applying it to diverse geographic regions. Our method offers a viable solution to the computational challenges of large-scale environmental monitoring.

Kaan Aydin, Joelle Hanna, Damian Borth

Earth observation satellites like Sentinel-1 (S1) and Sentinel-2 (S2) provide complementary remote sensing (RS) data, but S2 images are often unavailable due to cloud cover or data gaps. To address this, we propose a diffusion model (DM)-based approach for SAR-to-RGB translation, generating synthetic optical images from SAR inputs. We explore three different setups: two using Standard Diffusion, which reconstruct S2 images by adding and removing noise (one without and one with class conditioning), and one using Cold Diffusion, which blends S2 with S1 before removing the SAR signal. We evaluate the generated images in downstream tasks, including land cover classification and cloud removal. While generated images may not perfectly replicate real S2 data, they still provide valuable information. Our results show that class conditioning improves classification accuracy, while cloud removal performance remains competitive despite our approach not being optimized for it. Interestingly, despite exhibiting lower perceptual quality, the Cold Diffusion setup performs well in land cover classification, suggesting that traditional quantitative evaluation metrics may not fully reflect the practical utility of generated images. Our findings highlight the potential of DMs for SAR-to-RGB translation in RS applications where RGB images are missing.

Konstantin Schürholt, Michael W. Mahoney, Damian Borth

Learning representations of well-trained neural network models holds the promise to provide an understanding of the inner workings of those models. However, previous work has either faced limitations when processing larger networks or was task-specific to either discriminative or generative tasks. This paper introduces the SANE approach to weight-space learning. SANE overcomes previous limitations by learning task-agnostic representations of neural networks that are scalable to larger models of varying architectures and that show capabilities beyond a single task. Our method extends the idea of hyper-representations towards sequential processing of subsets of neural network weights, thus allowing one to embed larger neural networks as a set of tokens into the learned representation space. SANE reveals global model information from layer-wise embeddings, and it can sequentially generate unseen neural network models, which was unattainable with previous hyper-representation learning methods. Extensive empirical evaluation demonstrates that SANE matches or exceeds state-of-the-art performance on several weight representation learning benchmarks, particularly in initialization for new tasks and larger ResNet architectures.

Hamed Hemati, Damian Borth

The loss function plays an important role in optimizing the performance of a learning system. A crucial aspect of the loss function is the assignment of sample weights within a mini-batch during loss computation. In the context of continual learning (CL), most existing strategies uniformly treat samples when calculating the loss value, thereby assigning equal weights to each sample. While this approach can be effective in certain standard benchmarks, its optimal effectiveness, particularly in more complex scenarios, remains underexplored. This is particularly pertinent in training "in the wild," such as with self-training, where labeling is automated using a reference model. This paper introduces the Online Meta-learning for Sample Importance (OMSI) strategy that approximates sample weights for a mini-batch in an online CL stream using an inner- and meta-update mechanism. This is done by first estimating sample weight parameters for each sample in the mini-batch, then, updating the model with the adapted sample weights. We evaluate OMSI in two distinct experimental settings. First, we show that OMSI enhances both learning and retained accuracy in a controlled noisy-labeled data stream. Then, we test the strategy in three standard benchmarks and compare it with other popular replay-based strategies. This research aims to foster the ongoing exploration in the area of self-adaptive CL.

Timur Sattarov, Marco Schreyer, Damian Borth

The sharing of microdata, such as fund holdings and derivative instruments, by regulatory institutions presents a unique challenge due to strict data confidentiality and privacy regulations. These challenges often hinder the ability of both academics and practitioners to conduct collaborative research effectively. The emergence of generative models, particularly diffusion models, capable of synthesizing data mimicking the underlying distributions of real-world data presents a compelling solution. This work introduces 'FinDiff', a diffusion model designed to generate real-world financial tabular data for a variety of regulatory downstream tasks, for example economic scenario modeling, stress tests, and fraud detection. The model uses embedding encodings to model mixed modality financial data, comprising both categorical and numeric attributes. The performance of FinDiff in generating synthetic tabular financial data is evaluated against state-of-the-art baseline models using three real-world financial datasets (including two publicly available datasets and one proprietary dataset). Empirical results demonstrate that FinDiff excels in generating synthetic tabular financial data with high fidelity, privacy, and utility.

Hamed Hemati, Marco Schreyer, Damian Borth

International audit standards require the direct assessment of a financial statement's underlying accounting journal entries. Driven by advances in artificial intelligence, deep-learning inspired audit techniques emerged to examine vast quantities of journal entry data. However, in regular audits, most of the proposed methods are applied to learn from a comparably stationary journal entry population, e.g., of a financial quarter or year. Ignoring situations where audit relevant distribution changes are not evident in the training data or become incrementally available over time. In contrast, in continuous auditing, deep-learning models are continually trained on a stream of recorded journal entries, e.g., of the last hour. Resulting in situations where previous knowledge interferes with new information and will be entirely overwritten. This work proposes a continual anomaly detection framework to overcome both challenges and designed to learn from a stream of journal entry data experiences. The framework is evaluated based on deliberately designed audit scenarios and two real-world datasets. Our experimental results provide initial evidence that such a learning scheme offers the ability to reduce false-positive alerts and false-negative decisions.

Alex Bogun, Dimche Kostadinov, Damian Borth

Deep learning models have shown incredible performance on numerous image recognition, classification, and reconstruction tasks. Although very appealing and valuable due to their predictive capabilities, one common threat remains challenging to resolve. A specifically trained attacker can introduce malicious input perturbations to fool the network, thus causing potentially harmful mispredictions. Moreover, these attacks can succeed when the adversary has full access to the target model (white-box) and even when such access is limited (black-box setting). The ensemble of models can protect against such attacks but might be brittle under shared vulnerabilities in its members (attack transferability). To that end, this work proposes a novel diversity-promoting learning approach for the deep ensembles. The idea is to promote saliency map diversity (SMD) on ensemble members to prevent the attacker from targeting all ensemble members at once by introducing an additional term in our learning objective. During training, this helps us minimize the alignment between model saliencies to reduce shared member vulnerabilities and, thus, increase ensemble robustness to adversaries. We empirically show a reduced transferability between ensemble members and improved performance compared to the state-of-the-art ensemble defense against medium and high strength white-box attacks. In addition, we demonstrate that our approach combined with existing methods outperforms state-of-the-art ensemble algorithms for defense under white-box and black-box attacks.

Konstantin Schürholt, Dimche Kostadinov, Damian Borth

Self-Supervised Learning (SSL) has been shown to learn useful and information-preserving representations. Neural Networks (NNs) are widely applied, yet their weight space is still not fully understood. Therefore, we propose to use SSL to learn hyper-representations of the weights of populations of NNs. To that end, we introduce domain specific data augmentations and an adapted attention architecture. Our empirical evaluation demonstrates that self-supervised representation learning in this domain is able to recover diverse NN model characteristics. Further, we show that the proposed learned representations outperform prior work for predicting hyper-parameters, test accuracy, and generalization gap as well as transfer to out-of-distribution settings.

Shijun Wang, Dimche Kostadinov, Damian Borth

Voice Conversion (VC) for unseen speakers, also known as zero-shot VC, is an attractive research topic as it enables a range of applications like voice customizing, animation production, and others. Recent work in this area made progress with disentanglement methods that separate utterance content and speaker characteristics from speech audio recordings. However, many of these methods are subject to the leakage of prosody (e.g., pitch, volume), causing the speaker voice in the synthesized speech to be different from the desired target speakers. To prevent this issue, we propose a novel self-supervised approach that effectively learns disentangled pitch and volume representations that can represent the prosody styles of different speakers. We then use the learned prosodic representations as conditional information to train and enhance our VC model for zero-shot conversion. In our experiments, we show that our prosody representations are disentangled and rich in prosody information. Moreover, we demonstrate that the addition of our prosody representations improves our VC performance and surpasses state-of-the-art zero-shot VC performances.

Marco Schreyer, Timur Sattarov, Damian Borth

International audit standards require the direct assessment of a financial statement's underlying accounting transactions, referred to as journal entries. Recently, driven by the advances in artificial intelligence, deep learning inspired audit techniques have emerged in the field of auditing vast quantities of journal entry data. Nowadays, the majority of such methods rely on a set of specialized models, each trained for a particular audit task. At the same time, when conducting a financial statement audit, audit teams are confronted with (i) challenging time-budget constraints, (ii) extensive documentation obligations, and (iii) strict model interpretability requirements. As a result, auditors prefer to harness only a single preferably `multi-purpose' model throughout an audit engagement. We propose a contrastive self-supervised learning framework designed to learn audit task invariant accounting data representations to meet this requirement. The framework encompasses deliberate interacting data augmentation policies that utilize the attribute characteristics of journal entry data. We evaluate the framework on two real-world datasets of city payments and transfer the learned representations to three downstream audit tasks: anomaly detection, audit sampling, and audit documentation. Our experimental results provide empirical evidence that the proposed framework offers the ability to increase the efficiency of audits by learning rich and interpretable `multi-task' representations.

Tim Krappel, Alex Bogun, Damian Borth

Over the past years, topics ranging from climate change to human rights have seen increasing importance for investment decisions. Hence, investors (asset managers and asset owners) who wanted to incorporate these issues started to assess companies based on how they handle such topics. For this assessment, investors rely on specialized rating agencies that issue ratings along the environmental, social and governance (ESG) dimensions. Such ratings allow them to make investment decisions in favor of sustainability. However, rating agencies base their analysis on subjective assessment of sustainability reports, not provided by every company. Furthermore, due to human labor involved, rating agencies are currently facing the challenge to scale up the coverage in a timely manner. In order to alleviate these challenges and contribute to the overall goal of supporting sustainability, we propose a heterogeneous ensemble model to predict ESG ratings using fundamental data. This model is based on feedforward neural network, CatBoost and XGBoost ensemble members. Given the public availability of fundamental data, the proposed method would allow cost-efficient and scalable creation of initial ESG ratings (also for companies without sustainability reporting). Using our approach we are able to explain 54% of the variation in ratings R2 using fundamental data and outperform prior work in this area.

Saurabh Varshneya, Antoine Ledent, Robert A. Vandermeulen et al.

We propose a novel training methodology -- Concept Group Learning (CGL) -- that encourages training of interpretable CNN filters by partitioning filters in each layer into concept groups, each of which is trained to learn a single visual concept. We achieve this through a novel regularization strategy that forces filters in the same group to be active in similar image regions for a given layer. We additionally use a regularizer to encourage a sparse weighting of the concept groups in each layer so that a few concept groups can have greater importance than others. We quantitatively evaluate CGL's model interpretability using standard interpretability evaluation techniques and find that our method increases interpretability scores in most cases. Qualitatively we compare the image regions that are most active under filters learned using CGL versus filters learned without CGL and find that CGL activation regions more strongly concentrate around semantically relevant features.

Linus Scheibenreif, Michael Mommert, Damian Borth

Air pollution is a major driver of climate change. Anthropogenic emissions from the burning of fossil fuels for transportation and power generation emit large amounts of problematic air pollutants, including Greenhouse Gases (GHGs). Despite the importance of limiting GHG emissions to mitigate climate change, detailed information about the spatial and temporal distribution of GHG and other air pollutants is difficult to obtain. Existing models for surface-level air pollution rely on extensive land-use datasets which are often locally restricted and temporally static. This work proposes a deep learning approach for the prediction of ambient air pollution that only relies on remote sensing data that is globally available and frequently updated. Combining optical satellite imagery with satellite-based atmospheric column density air pollution measurements enables the scaling of air pollution estimates (in this case NO$_2$) to high spatial resolution (up to $\sim$10m) at arbitrary locations and adds a temporal component to these estimates. The proposed model performs with high accuracy when evaluated against air quality measurements from ground stations (mean absolute error $<$6$~μg/m^3$). Our results enable the identification and temporal monitoring of major sources of air pollution and GHGs.

Michael Mommert, Linus Scheibenreif, Joëlle Hanna et al.

Satellite remote imaging enables the detailed study of land use patterns on a global scale. We investigate the possibility to improve the information content of traditional land use classification by identifying the nature of industrial sites from medium-resolution remote sensing images. In this work, we focus on classifying different types of power plants from Sentinel-2 imaging data. Using a ResNet-50 deep learning model, we are able to achieve a mean accuracy of 90.0% in distinguishing 10 different power plant types and a background class. Furthermore, we are able to identify the cooling mechanisms utilized in thermal power plants with a mean accuracy of 87.5%. Our results enable us to qualitatively investigate the energy mix from Sentinel-2 imaging data, and prove the feasibility to classify industrial sites on a global scale from freely available satellite imagery.

Shijun Wang, Damian Borth

Voice conversion (VC) is a task that transforms voice from target audio to source without losing linguistic contents, it is challenging especially when source and target speakers are unseen during training (zero-shot VC). Previous approaches require a pre-trained model or linguistic data to do the zero-shot conversion. Meanwhile, VC models with Vector Quantization (VQ) or Instance Normalization (IN) are able to disentangle contents from audios and achieve successful conversions. However, disentanglement in these models highly relies on heavily constrained bottleneck layers, thus, the sound quality is drastically sacrificed. In this paper, we propose NoiseVC, an approach that can disentangle contents based on VQ and Contrastive Predictive Coding (CPC). Additionally, Noise Augmentation is performed to further enhance disentanglement capability. We conduct several experiments and demonstrate that NoiseVC has a strong disentanglement ability with a small sacrifice of quality.

Hamed Hemati, Damian Borth

Training a multi-speaker Text-to-Speech (TTS) model from scratch is computationally expensive and adding new speakers to the dataset requires the model to be re-trained. The naive solution of sequential fine-tuning of a model for new speakers can lead to poor performance of older speakers. This phenomenon is known as catastrophic forgetting. In this paper, we look at TTS modeling from a continual learning perspective, where the goal is to add new speakers without forgetting previous speakers. Therefore, we first propose an experimental setup and show that serial fine-tuning for new speakers can cause the forgetting of the earlier speakers. Then we exploit two well-known techniques for continual learning, namely experience replay and weight regularization. We reveal how one can mitigate the effect of degradation in speech synthesis diversity in sequential training of new speakers using these methods. Finally, we present a simple extension to experience replay to improve the results in extreme setups where we have access to very small buffers.

Marco Schreyer, Chistian Schulze, Damian Borth

Nowadays, organizations collect vast quantities of sensitive information in `Enterprise Resource Planning' (ERP) systems, such as accounting relevant transactions, customer master data, or strategic sales price information. The leakage of such information poses a severe threat for companies as the number of incidents and the reputational damage to those experiencing them continue to increase. At the same time, discoveries in deep learning research revealed that machine learning models could be maliciously misused to create new attack vectors. Understanding the nature of such attacks becomes increasingly important for the (internal) audit and fraud examination practice. The creation of such an awareness holds in particular for the fraudulent data leakage using deep learning-based steganographic techniques that might remain undetected by state-of-the-art `Computer Assisted Audit Techniques' (CAATs). In this work, we introduce a real-world `threat model' designed to leak sensitive accounting data. In addition, we show that a deep steganographic process, constituted by three neural networks, can be trained to hide such data in unobtrusive `day-to-day' images. Finally, we provide qualitative and quantitative evaluations on two publicly available real-world payment datasets.

Michael Mommert, Mario Sigel, Marcel Neuhausler et al.

The major driver of global warming has been identified as the anthropogenic release of greenhouse gas (GHG) emissions from industrial activities. The quantitative monitoring of these emissions is mandatory to fully understand their effect on the Earth's climate and to enforce emission regulations on a large scale. In this work, we investigate the possibility to detect and quantify industrial smoke plumes from globally and freely available multi-band image data from ESA's Sentinel-2 satellites. Using a modified ResNet-50, we can detect smoke plumes of different sizes with an accuracy of 94.3%. The model correctly ignores natural clouds and focuses on those imaging channels that are related to the spectral absorption from aerosols and water vapor, enabling the localization of smoke. We exploit this localization ability and train a U-Net segmentation model on a labeled sub-sample of our data, resulting in an Intersection-over-Union (IoU) metric of 0.608 and an overall accuracy for the detection of any smoke plume of 94.0%; on average, our model can reproduce the area covered by smoke in an image to within 5.6%. The performance of our model is mostly limited by occasional confusion with surface objects, the inability to identify semi-transparent smoke, and human limitations to properly identify smoke based on RGB-only images. Nevertheless, our results enable us to reliably detect and qualitatively estimate the level of smoke activity in order to monitor activity in industrial plants across the globe. Our data set and code base are publicly available.

Hamed Hemati, Damian Borth

Recent neural Text-to-Speech (TTS) models have been shown to perform very well when enough data is available. However, fine-tuning them for new speakers or languages is not straightforward in a low-resource setup. In this paper, we show that by applying minor modifications to a Tacotron model, one can transfer an existing TTS model for new speakers from the same or a different language using only 20 minutes of data. For this purpose, we first introduce a base multi-lingual Tacotron with language-agnostic input, then demonstrate how transfer learning is done for different scenarios of speaker adaptation without exploiting any pre-trained speaker encoder or code-switching technique. We evaluate the transferred model in both subjective and objective ways.