Federico Raue

Tobias Christian Nauen, Stanislav Frolov, Federico Raue et al.

CutMix has become the de facto standard mixing augmentation, yet its label assignment rests on a flawed assumption: The area of the pasted patch faithfully reflects its semantic contribution to the mixed image. In practice, however, patches frequently land on background regions, assigning label credit to classes whose objects are not visible. The mean discrepancy of the CutMix label and the semantic object area is $21.5\%$. In $17\%$ of samples an image contributes zero visible object pixels yet receives nonzero label weight. We propose Object-Aware CutMix (OA-CutMix), which corrects this bias by replacing the area-based CutMix weight with one derived from precomputed segmentation masks, assigning labels in proportion to the visible object area each image contributes to the mix. The image mixing procedure is left entirely unchanged. We evaluate OA-CutMix against 10+ static and dynamic mixing methods across 4 architectures and 6 datasets. OA-CutMix consistently achieves the highest accuracy over all tasks, outperforming even dynamic mixing methods, but at a fraction of the training-time cost. Improvements are largest for small objects, where the label bias from CutMix is greatest. Thus, correcting the label is sufficient to match or exceed the performance of methods modifying the image mixing algorithm.

Brian Moser, Federico Raue, Stanislav Frolov et al.

With the advent of Deep Learning (DL), Super-Resolution (SR) has also become a thriving research area. However, despite promising results, the field still faces challenges that require further research e.g., allowing flexible upsampling, more effective loss functions, and better evaluation metrics. We review the domain of SR in light of recent advances, and examine state-of-the-art models such as diffusion (DDPM) and transformer-based SR models. We present a critical discussion on contemporary strategies used in SR, and identify promising yet unexplored research directions. We complement previous surveys by incorporating the latest developments in the field such as uncertainty-driven losses, wavelet networks, neural architecture search, novel normalization methods, and the latests evaluation techniques. We also include several visualizations for the models and methods throughout each chapter in order to facilitate a global understanding of the trends in the field. This review is ultimately aimed at helping researchers to push the boundaries of DL applied to SR.

Brian Moser, Stanislav Frolov, Federico Raue et al.

This paper presents a novel Diffusion-Wavelet (DiWa) approach for Single-Image Super-Resolution (SISR). It leverages the strengths of Denoising Diffusion Probabilistic Models (DDPMs) and Discrete Wavelet Transformation (DWT). By enabling DDPMs to operate in the DWT domain, our DDPM models effectively hallucinate high-frequency information for super-resolved images on the wavelet spectrum, resulting in high-quality and detailed reconstructions in image space. Quantitatively, we outperform state-of-the-art diffusion-based SISR methods, namely SR3 and SRDiff, regarding PSNR, SSIM, and LPIPS on both face (8x scaling) and general (4x scaling) SR benchmarks. Meanwhile, using DWT enabled us to use fewer parameters than the compared models: 92M parameters instead of 550M compared to SR3 and 9.3M instead of 12M compared to SRDiff. Additionally, our method outperforms other state-of-the-art generative methods on classical general SR datasets while saving inference time. Finally, our work highlights its potential for various applications.

Brian Moser, Federico Raue, Jörn Hees et al.

Neural Architecture Search (NAS) defines the design of Neural Networks as a search problem. Unfortunately, NAS is computationally intensive because of various possibilities depending on the number of elements in the design and the possible connections between them. In this work, we extensively analyze the role of the dataset size based on several sampling approaches for reducing the dataset size (unsupervised and supervised cases) as an agnostic approach to reduce search time. We compared these techniques with four common NAS approaches in NAS-Bench-201 in roughly 1,400 experiments on CIFAR-100. One of our surprising findings is that in most cases we can reduce the amount of training data to 25\%, consequently reducing search time to 25\%, while at the same time maintaining the same accuracy as if training on the full dataset. Additionally, some designs derived from subsets out-perform designs derived from the full dataset by up to 22 p.p. accuracy.

Brian B. Moser, Stanislav Frolov, Federico Raue et al.

Diffusion models in image Super-Resolution (SR) treat all image regions uniformly, which risks compromising the overall image quality by potentially introducing artifacts during denoising of less-complex regions. To address this, we propose ``You Only Diffuse Areas'' (YODA), a dynamic attention-guided diffusion process for image SR. YODA selectively focuses on spatial regions defined by attention maps derived from the low-resolution images and the current denoising time step. This time-dependent targeting enables a more efficient conversion to high-resolution outputs by focusing on areas that benefit the most from the iterative refinement process, i.e., detail-rich objects. We empirically validate YODA by extending leading diffusion-based methods SR3, DiffBIR, and SRDiff. Our experiments demonstrate new state-of-the-art performances in face and general SR tasks across PSNR, SSIM, and LPIPS metrics. As a side effect, we find that YODA reduces color shift issues and stabilizes training with small batches.

Tobias Christian Nauen, Sebastian Palacio, Federico Raue et al.

Self-attention in Transformers comes with a high computational cost because of their quadratic computational complexity, but their effectiveness in addressing problems in language and vision has sparked extensive research aimed at enhancing their efficiency. However, diverse experimental conditions, spanning multiple input domains, prevent a fair comparison based solely on reported results, posing challenges for model selection. To address this gap in comparability, we perform a large-scale benchmark of more than 45 models for image classification, evaluating key efficiency aspects, including accuracy, speed, and memory usage. Our benchmark provides a standardized baseline for efficiency-oriented transformers. We analyze the results based on the Pareto front -- the boundary of optimal models. Surprisingly, despite claims of other models being more efficient, ViT remains Pareto optimal across multiple metrics. We observe that hybrid attention-CNN models exhibit remarkable inference memory- and parameter-efficiency. Moreover, our benchmark shows that using a larger model in general is more efficient than using higher resolution images. Thanks to our holistic evaluation, we provide a centralized resource for practitioners and researchers, facilitating informed decisions when selecting or developing efficient transformers.

Brian B. Moser, Stanislav Frolov, Federico Raue et al.

This work introduces Differential Wavelet Amplifier (DWA), a drop-in module for wavelet-based image Super-Resolution (SR). DWA invigorates an approach recently receiving less attention, namely Discrete Wavelet Transformation (DWT). DWT enables an efficient image representation for SR and reduces the spatial area of its input by a factor of 4, the overall model size, and computation cost, framing it as an attractive approach for sustainable ML. Our proposed DWA model improves wavelet-based SR models by leveraging the difference between two convolutional filters to refine relevant feature extraction in the wavelet domain, emphasizing local contrasts and suppressing common noise in the input signals. We show its effectiveness by integrating it into existing SR models, e.g., DWSR and MWCNN, and demonstrate a clear improvement in classical SR tasks. Moreover, DWA enables a direct application of DWSR and MWCNN to input image space, reducing the DWT representation channel-wise since it omits traditional DWT.

Brian Moser, Federico Raue, Jörn Hees et al.

We present new Recurrent Neural Network (RNN) cells for image classification using a Neural Architecture Search (NAS) approach called DARTS. We are interested in the ReNet architecture, which is a RNN based approach presented as an alternative for convolutional and pooling steps. ReNet can be defined using any standard RNN cells, such as LSTM and GRU. One limitation is that standard RNN cells were designed for one dimensional sequential data and not for two dimensions like it is the case for image classification. We overcome this limitation by using DARTS to find new cell designs. We compare our results with ReNet that uses GRU and LSTM cells. Our found cells outperform the standard RNN cells on CIFAR-10 and SVHN. The improvements on SVHN indicate generalizability, as we derived the RNN cell designs from CIFAR-10 without performing a new cell search for SVHN.

Ahmed Anwar, Brian Moser, Dayananda Herurkar et al.

The emergence of federated learning (FL) presents a promising approach to leverage decentralized data while preserving privacy. Furthermore, the combination of FL and anomaly detection is particularly compelling because it allows for detecting rare and critical anomalies (usually also rare in locally gathered data) in sensitive data from multiple sources, such as cybersecurity and healthcare. However, benchmarking the performance of anomaly detection methods in FL environments remains an underexplored area. This paper introduces FedAD-Bench, a unified benchmark for evaluating unsupervised anomaly detection algorithms within the context of FL. We systematically analyze and compare the performance of recent deep learning anomaly detection models under federated settings, which were typically assessed solely in centralized settings. FedAD-Bench encompasses diverse datasets and metrics to provide a holistic evaluation. Through extensive experiments, we identify key challenges such as model aggregation inefficiencies and metric unreliability. We present insights into FL's regularization effects, revealing scenarios in which it outperforms centralized approaches due to its inherent ability to mitigate overfitting. Our work aims to establish a standardized benchmark to guide future research and development in federated anomaly detection, promoting reproducibility and fair comparison across studies.

Tobias Christian Nauen, Stanislav Frolov, Brian Bernhard Moser et al.

The platonic representation hypothesis suggests that sufficiently large models converge to a shared representation geometry, even across modalities. Motivated by this, we ask: Can the semantic knowledge of a language model efficiently improve a vision model? As an answer, we introduce TextTeacher, a simple auxiliary objective that injects text embeddings as additional information into image classification training. TextTeacher uses readily available image captions, a pre-trained and frozen text encoder, and a lightweight projection to produce semantic anchors that efficiently guide representations during training while leaving the inference-time model unchanged. On ImageNet with standard ViT backbones, TextTeacher improves accuracy by up to +2.7 percentage points (p.p.) and yields consistent transfer gains (on average +1.0 p.p.) under the same recipe and compute. It outperforms vision knowledge distillation, yielding more accuracy at a constant compute budget or similar accuracy, but 33% faster. Our analysis indicates that TextTeacher acts as a feature-space preconditioner, shaping deeper layers in the first stages of training, and aiding generalization by supplying complementary semantic cues. TextTeacher adds negligible overhead, requires no costly multimodal training of the target model and preserves the simplicity and latency of pure vision models. Project page with code and captions: https://nauen-it.de/publications/text-teacher

Brian B. Moser, Federico Raue, Sebastian Palacio et al.

Dataset distillation seeks to condense datasets into smaller but highly representative synthetic samples. While diffusion models now lead all generative benchmarks, current distillation methods avoid them and rely instead on GANs or autoencoders, or, at best, sampling from a fixed diffusion prior. This trend arises because naive backpropagation through the long denoising chain leads to vanishing gradients, which prevents effective synthetic sample optimization. To address this limitation, we introduce Latent Dataset Distillation with Diffusion Models (LD3M), the first method to learn gradient-based distilled latents and class embeddings end-to-end through a pre-trained latent diffusion model. A linearly decaying skip connection, injected from the initial noisy state into every reverse step, preserves the gradient signal across dozens of timesteps without requiring diffusion weight fine-tuning. Across multiple ImageNet subsets at 128x128 and 256x256, LD3M improves downstream accuracy by up to 4.8 percentage points (1 IPC) and 4.2 points (10 IPC) over the prior state-of-the-art. The code for LD3M is provided at https://github.com/Brian-Moser/prune_and_distill.

Krzysztof Adamkiewicz, Brian Moser, Stanislav Frolov et al.

Recent text-to-image (T2I) diffusion models produce visually stunning images and demonstrate excellent prompt following. But do they perform well as synthetic vision data generators? In this work, we revisit the promise of synthetic data as a scalable substitute for real training sets and uncover a surprising performance regression. We generate large-scale synthetic datasets using state-of-the-art T2I models released between 2022 and 2025, train standard classifiers solely on this synthetic data, and evaluate them on real test data. Despite observable advances in visual fidelity and prompt adherence, classification accuracy on real test data consistently declines with newer T2I models as training data generators. Our analysis reveals a hidden trend: These models collapse to a narrow, aesthetic-centric distribution that undermines diversity and label-image alignment. Overall, our findings challenge a growing assumption in vision research, namely that progress in generative realism implies progress in data realism. We thus highlight an urgent need to rethink the capabilities of modern T2I models as reliable training data generators.

Brian B. Moser, Shalini Strode, Federico Raue et al.

Dataset distillation (DD) promises compact yet faithful synthetic data, but existing approaches often inherit the inductive bias of a single teacher model. As dataset size increases, this bias drives generation toward overly smooth, homogeneous samples, reducing intra-class diversity and limiting generalization. We present PRISM (PRIors from diverse Source Models), a framework that disentangles architectural priors during synthesis. PRISM decouples the logit-matching and regularization objectives, supervising them with different teacher architectures: a primary model for logits and a stochastic subset for batch-normalization (BN) alignment. On ImageNet-1K, PRISM consistently and reproducibly outperforms single-teacher methods (e.g., SRe2L) and recent multi-teacher variants (e.g., G-VBSM) at low- and mid-IPC regimes. The generated data also show significantly richer intra-class diversity, as reflected by a notable drop in cosine similarity between features. We further analyze teacher selection strategies (pre- vs. intra-distillation) and introduce a scalable cross-class batch formation scheme for fast parallel synthesis. Code will be released after the review period.

Tobias Christian Nauen, Brian Moser, Federico Raue et al.

Transformers, particularly Vision Transformers (ViTs), have achieved state-of-the-art performance in large-scale image classification. However, they often require large amounts of data and can exhibit biases that limit their robustness and generalizability. This paper introduces ForAug, a novel data augmentation scheme that addresses these challenges and explicitly includes inductive biases, which commonly are part of the neural network architecture, into the training data. ForAug is constructed by using pretrained foundation models to separate and recombine foreground objects with different backgrounds, enabling fine-grained control over image composition during training. It thus increases the data diversity and effective number of training samples. We demonstrate that training on ForNet, the application of ForAug to ImageNet, significantly improves the accuracy of ViTs and other architectures by up to 4.5 percentage points (p.p.) on ImageNet and 7.3 p.p. on downstream tasks. Importantly, ForAug enables novel ways of analyzing model behavior and quantifying biases. Namely, we introduce metrics for background robustness, foreground focus, center bias, and size bias and show that training on ForNet substantially reduces these biases compared to training on ImageNet. In summary, ForAug provides a valuable tool for analyzing and mitigating biases, enabling the development of more robust and reliable computer vision models. Our code and dataset are publicly available at https://github.com/tobna/ForAug.

Stanislav Frolov, Avneesh Sharma, Jörn Hees et al.

Conditional image synthesis from layout has recently attracted much interest. Previous approaches condition the generator on object locations as well as class labels but lack fine-grained control over the diverse appearance aspects of individual objects. Gaining control over the image generation process is fundamental to build practical applications with a user-friendly interface. In this paper, we propose a method for attribute controlled image synthesis from layout which allows to specify the appearance of individual objects without affecting the rest of the image. We extend a state-of-the-art approach for layout-to-image generation to additionally condition individual objects on attributes. We create and experiment on a synthetic, as well as the challenging Visual Genome dataset. Our qualitative and quantitative results show that our method can successfully control the fine-grained details of individual objects when modelling complex scenes with multiple objects. Source code, dataset and pre-trained models are publicly available (https://github.com/stanifrolov/AttrLostGAN).

Brian B. Moser, Arundhati S. Shanbhag, Federico Raue et al.

Diffusion Models (DMs) have disrupted the image Super-Resolution (SR) field and further closed the gap between image quality and human perceptual preferences. They are easy to train and can produce very high-quality samples that exceed the realism of those produced by previous generative methods. Despite their promising results, they also come with new challenges that need further research: high computational demands, comparability, lack of explainability, color shifts, and more. Unfortunately, entry into this field is overwhelming because of the abundance of publications. To address this, we provide a unified recount of the theoretical foundations underlying DMs applied to image SR and offer a detailed analysis that underscores the unique characteristics and methodologies within this domain, distinct from broader existing reviews in the field. This survey articulates a cohesive understanding of DM principles and explores current research avenues, including alternative input domains, conditioning techniques, guidance mechanisms, corruption spaces, and zero-shot learning approaches. By offering a detailed examination of the evolution and current trends in image SR through the lens of DMs, this survey sheds light on the existing challenges and charts potential future directions, aiming to inspire further innovation in this rapidly advancing area.

Shalini Sarode, Brian Moser, Joachim Folz et al.

The Forward-Forward (FF) algorithm presents a compelling, bio-inspired alternative to backpropagation. However, while efficient in training, it has a computationally prohibitive inference process that requires a separate forward pass for every class that is evaluated. In this work, we introduce the Hyperspherical Forward-Forward (HFF), a novel reformulation that resolves this critical bottleneck. Our core innovation is to reframe the local objective of each layer from a binary goodness-of-fit task to a direct multi-class classification problem within a hyperspherical feature space. We achieve this by learning a set of class-specific, unit-norm prototypes that act as geometric anchors and implicit negatives. This architectural innovation preserves the benefits of local training while enabling weight update and inference in a single forward pass, making it >40x faster than the original FF algorithm. Our method is simple to implement, scales effectively to modern convolutional architectures, and achieves superior accuracy on standard image classification benchmarks, closing the gap with backpropagation. Most notably, we are among the first greedy local-learning methods to report over 25% top-1 accuracy on ImageNet-1k, and 65.96% with transfer learning.

Brian B. Moser, Federico Raue, Andreas Dengel

In image Super-Resolution (SR), relying on large datasets for training is a double-edged sword. While offering rich training material, they also demand substantial computational and storage resources. In this work, we analyze dataset pruning to solve these challenges. We introduce a novel approach that reduces a dataset to a core-set of training samples, selected based on their loss values as determined by a simple pre-trained SR model. By focusing the training on just 50\% of the original dataset, specifically on the samples characterized by the highest loss values, we achieve results comparable to or surpassing those obtained from training on the entire dataset. Interestingly, our analysis reveals that the top 5\% of samples with the highest loss values negatively affect the training process. Excluding these samples and adjusting the selection to favor easier samples further enhances training outcomes. Our work opens new perspectives to the untapped potential of dataset pruning in image SR. It suggests that careful selection of training data based on loss-value metrics can lead to better SR models, challenging the conventional wisdom that more data inevitably leads to better performance.

Brian B. Moser, Arundhati S. Shanbhag, Stanislav Frolov et al.

Coreset selection targets the challenge of finding a small, representative subset of a large dataset that preserves essential patterns for effective machine learning. Although several surveys have examined data reduction strategies before, most focus narrowly on either classical geometry-based methods or active learning techniques. In contrast, this survey presents a more comprehensive view by unifying three major lines of coreset research, namely, training-free, training-oriented, and label-free approaches, into a single taxonomy. We present subfields often overlooked by existing work, including submodular formulations, bilevel optimization, and recent progress in pseudo-labeling for unlabeled datasets. Additionally, we examine how pruning strategies influence generalization and neural scaling laws, offering new insights that are absent from prior reviews. Finally, we compare these methods under varying computational, robustness, and performance demands and highlight open challenges, such as robustness, outlier filtering, and adapting coreset selection to foundation models, for future research.

Brian B. Moser, Federico Raue, Tobias C. Nauen et al.

Dataset distillation has gained significant interest in recent years, yet existing approaches typically distill from the entire dataset, potentially including non-beneficial samples. We introduce a novel "Prune First, Distill After" framework that systematically prunes datasets via loss-based sampling prior to distillation. By leveraging pruning before classical distillation techniques and generative priors, we create a representative core-set that leads to enhanced generalization for unseen architectures - a significant challenge of current distillation methods. More specifically, our proposed framework significantly boosts distilled quality, achieving up to a 5.2 percentage points accuracy increase even with substantial dataset pruning, i.e., removing 80% of the original dataset prior to distillation. Overall, our experimental results highlight the advantages of our easy-sample prioritization and cross-architecture robustness, paving the way for more effective and high-quality dataset distillation.

Brian B. Moser, Ahmed Anwar, Federico Raue et al.

Traditional blind image SR methods need to model real-world degradations precisely. Consequently, current research struggles with this dilemma by assuming idealized degradations, which leads to limited applicability to actual user data. Moreover, the ideal scenario - training models on data from the targeted user base - presents significant privacy concerns. To address both challenges, we propose to fuse image SR with federated learning, allowing real-world degradations to be directly learned from users without invading their privacy. Furthermore, it enables optimization across many devices without data centralization. As this fusion is underexplored, we introduce new benchmarks specifically designed to evaluate new SR methods in this federated setting. By doing so, we employ known degradation modeling techniques from SR research. However, rather than aiming to mirror real degradations, our benchmarks use these degradation models to simulate the variety of degradations found across clients within a distributed user base. This distinction is crucial as it circumvents the need to precisely model real-world degradations, which limits contemporary blind image SR research. Our proposed benchmarks investigate blind image SR under new aspects, namely differently distributed degradation types among users and varying user numbers. We believe new methods tested within these benchmarks will perform more similarly in an application, as the simulated scenario addresses the variety while federated learning enables the training on actual degradations.

Tobias Dietz, Brian B. Moser, Tobias Nauen et al.

Dataset distillation aims to compress large datasets into compact yet highly informative subsets that preserve the training behavior of the original data. While this concept has gained traction in classification, its potential for image Super-Resolution (SR) remains largely untapped. In this work, we conduct the first systematic study of dataset distillation for SR, evaluating both pixel- and latent-space formulations. We show that a distilled dataset, occupying only 8.88% of the original size, can train SR models that retain nearly the same reconstruction fidelity as those trained on full datasets. Furthermore, we analyze how initialization strategies and distillation objectives affect efficiency, convergence, and visual quality. Our findings highlight the feasibility of SR dataset distillation and establish foundational insights for memory- and compute-efficient generative restoration models.

Arundhati S. Shanbhag, Brian B. Moser, Tobias C. Nauen et al.

Diffusion models, celebrated for their generative capabilities, have recently demonstrated surprising effectiveness in image classification tasks by using Bayes' theorem. Yet, current diffusion classifiers must evaluate every label candidate for each input, creating high computational costs that impede their use in large-scale applications. To address this limitation, we propose a Hierarchical Diffusion Classifier (HDC) that exploits hierarchical label structures or well-defined parent-child relationships in the dataset. By pruning irrelevant high-level categories and refining predictions only within relevant subcategories (leaf nodes and sub-trees), HDC reduces the total number of class evaluations. As a result, HDC can speed up inference by as much as 60% while preserving and sometimes even improving classification accuracy. In summary, our work provides a tunable control mechanism between speed and precision, making diffusion-based classification more feasible for large-scale applications.

Brian B. Moser, Stanislav Frolov, Tobias C. Nauen et al.

Large-scale, pre-trained Text-to-Image (T2I) diffusion models have gained significant popularity in image generation tasks and have shown unexpected potential in image Super-Resolution (SR). However, most existing T2I diffusion models are trained with a resolution limit of 512x512, making scaling beyond this resolution an unresolved but necessary challenge for image SR. In this work, we introduce a novel approach that, for the first time, enables these models to generate 2K, 4K, and even 8K images without any additional training. Our method leverages MultiDiffusion, which distributes the generation across multiple diffusion paths to ensure global coherence at larger scales, and local degradation-aware prompt extraction, which guides the T2I model to reconstruct fine local structures according to its low-resolution input. These innovations unlock higher resolutions, allowing T2I diffusion models to be applied to image SR tasks without limitation on resolution.

Brian B. Moser, Tobias C. Nauen, Arundhati S. Shanbhag et al.

The goal of coreset selection is to identify representative subsets of datasets for efficient model training. Yet, existing approaches paradoxically require expensive training-based signals, e.g., gradients, decision boundary estimates or forgetting counts, computed over the entire dataset prior to pruning, which undermines their very purpose by requiring training on samples they aim to avoid. We introduce SubZeroCore, a novel, training-free coreset selection method that integrates submodular coverage and density into a single, unified objective. To achieve this, we introduce a sampling strategy based on a closed-form solution to optimally balance these objectives, guided by a single hyperparameter that explicitly controls the desired coverage for local density measures. Despite no training, extensive evaluations show that SubZeroCore matches training-based baselines and significantly outperforms them at high pruning rates, while dramatically reducing computational overhead. SubZeroCore also demonstrates superior robustness to label noise, highlighting its practical effectiveness and scalability for real-world scenarios.

Brian B. Moser, Arundhati S. Shanbhag, Tobias C. Nauen et al.

The goal of coreset selection methods is to identify representative subsets of datasets for efficient model training. Yet, existing methods often ignore the possibility of annotation errors and require fixed pruning ratios, making them impractical in real-world settings. We present HyperCore, a robust and adaptive coreset selection framework designed explicitly for noisy environments. HyperCore leverages lightweight hypersphere models learned per class, embedding in-class samples close to a hypersphere center while naturally segregating out-of-class samples based on their distance. By using Youden's J statistic, HyperCore can adaptively select pruning thresholds, enabling automatic, noise-aware data pruning without hyperparameter tuning. Our experiments reveal that HyperCore consistently surpasses state-of-the-art coreset selection methods, especially under noisy and low-data regimes. HyperCore effectively discards mislabeled and ambiguous points, yielding compact yet highly informative subsets suitable for scalable and noise-free learning.

Sanath Budakegowdanadoddi Nagaraju, Brian Bernhard Moser, Tobias Christian Nauen et al.

Transformer-based architectures have recently advanced the image reconstruction quality of super-resolution (SR) models. Yet, their scalability remains limited by quadratic attention costs and coarse patch embeddings that weaken pixel-level fidelity. We propose TaylorIR, a plug-and-play framework that enforces 1x1 patch embeddings for true pixel-wise reasoning and replaces conventional self-attention with TaylorShift, a Taylor-series-based attention mechanism enabling full token interactions with near-linear complexity. Across multiple SR benchmarks, TaylorIR delivers state-of-the-art performance while reducing memory consumption by up to 60%, effectively bridging the gap between fine-grained detail restoration and efficient transformer scaling.

Fatemeh Azimi, Jean-Francois Jacques Nicolas Nies, Sebastian Palacio et al.

Curriculum learning is a bio-inspired training technique that is widely adopted to machine learning for improved optimization and better training of neural networks regarding the convergence rate or obtained accuracy. The main concept in curriculum learning is to start the training with simpler tasks and gradually increase the level of difficulty. Therefore, a natural question is how to determine or generate these simpler tasks. In this work, we take inspiration from Spatial Transformer Networks (STNs) in order to form an easy-to-hard curriculum. As STNs have been proven to be capable of removing the clutter from the input images and obtaining higher accuracy in image classification tasks, we hypothesize that images processed by STNs can be seen as easier tasks and utilized in the interest of curriculum learning. To this end, we study multiple strategies developed for shaping the training curriculum, using the data generated by STNs. We perform various experiments on cluttered MNIST and Fashion-MNIST datasets, where on the former, we obtain an improvement of $3.8$pp in classification accuracy compared to the baseline.

Fatemeh Azimi, Federico Raue, Joern Hees et al.

Spatial Transformer Networks (STN) can generate geometric transformations which modify input images to improve the classifier's performance. In this work, we combine the idea of STN with Reinforcement Learning (RL). To this end, we break the affine transformation down into a sequence of simple and discrete transformations. We formulate the task as a Markovian Decision Process (MDP) and use RL to solve this sequential decision-making problem. STN architectures learn the transformation parameters by minimizing the classification error and backpropagating the gradients through a sub-differentiable sampling module. In our method, we are not bound to the differentiability of the sampling modules. Moreover, we have freedom in designing the objective rather than only minimizing the error; e.g., we can directly set the target as maximizing the accuracy. We design multiple experiments to verify the effectiveness of our method using cluttered MNIST and Fashion-MNIST datasets and show that our method outperforms STN with a proper definition of MDP components.

Andrey Guzhov, Federico Raue, Jörn Hees et al.

In the past, the rapidly evolving field of sound classification greatly benefited from the application of methods from other domains. Today, we observe the trend to fuse domain-specific tasks and approaches together, which provides the community with new outstanding models. In this work, we present an extension of the CLIP model that handles audio in addition to text and images. Our proposed model incorporates the ESResNeXt audio-model into the CLIP framework using the AudioSet dataset. Such a combination enables the proposed model to perform bimodal and unimodal classification and querying, while keeping CLIP's ability to generalize to unseen datasets in a zero-shot inference fashion. AudioCLIP achieves new state-of-the-art results in the Environmental Sound Classification (ESC) task, out-performing other approaches by reaching accuracies of 90.07% on the UrbanSound8K and 97.15% on the ESC-50 datasets. Further it sets new baselines in the zero-shot ESC-task on the same datasets (68.78% and 69.40%, respectively). Finally, we also assess the cross-modal querying performance of the proposed model as well as the influence of full and partial training on the results. For the sake of reproducibility, our code is published.

Ricard Durall, Stanislav Frolov, Jörn Hees et al.

Transformer models have recently attracted much interest from computer vision researchers and have since been successfully employed for several problems traditionally addressed with convolutional neural networks. At the same time, image synthesis using generative adversarial networks (GANs) has drastically improved over the last few years. The recently proposed TransGAN is the first GAN using only transformer-based architectures and achieves competitive results when compared to convolutional GANs. However, since transformers are data-hungry architectures, TransGAN requires data augmentation, an auxiliary super-resolution task during training, and a masking prior to guide the self-attention mechanism. In this paper, we study the combination of a transformer-based generator and convolutional discriminator and successfully remove the need of the aforementioned required design choices. We evaluate our approach by conducting a benchmark of well-known CNN discriminators, ablate the size of the transformer-based generator, and show that combining both architectural elements into a hybrid model leads to better results. Furthermore, we investigate the frequency spectrum properties of generated images and observe that our model retains the benefits of an attention based generator.

Andrey Guzhov, Federico Raue, Jörn Hees et al.

Environmental Sound Classification (ESC) is a rapidly evolving field that recently demonstrated the advantages of application of visual domain techniques to the audio-related tasks. Previous studies indicate that the domain-specific modification of cross-domain approaches show a promise in pushing the whole area of ESC forward. In this paper, we present a new time-frequency transformation layer that is based on complex frequency B-spline (fbsp) wavelets. Being used with a high-performance audio classification model, the proposed fbsp-layer provides an accuracy improvement over the previously used Short-Time Fourier Transform (STFT) on standard datasets. We also investigate the influence of different pre-training strategies, including the joint use of two large-scale datasets for weight initialization: ImageNet and AudioSet. Our proposed model out-performs other approaches by achieving accuracies of 95.20 % on the ESC-50 and 89.14 % on the UrbanSound8K datasets. Additionally, we assess the increase of model robustness against additive white Gaussian noise and reduction of an effective sample rate introduced by the proposed layer and demonstrate that the fbsp-layer improves the model's ability to withstand signal perturbations, in comparison to STFT-based training. For the sake of reproducibility, our code is made available.

Stanislav Frolov, Tobias Hinz, Federico Raue et al.

With the advent of generative adversarial networks, synthesizing images from textual descriptions has recently become an active research area. It is a flexible and intuitive way for conditional image generation with significant progress in the last years regarding visual realism, diversity, and semantic alignment. However, the field still faces several challenges that require further research efforts such as enabling the generation of high-resolution images with multiple objects, and developing suitable and reliable evaluation metrics that correlate with human judgement. In this review, we contextualize the state of the art of adversarial text-to-image synthesis models, their development since their inception five years ago, and propose a taxonomy based on the level of supervision. We critically examine current strategies to evaluate text-to-image synthesis models, highlight shortcomings, and identify new areas of research, ranging from the development of better datasets and evaluation metrics to possible improvements in architectural design and model training. This review complements previous surveys on generative adversarial networks with a focus on text-to-image synthesis which we believe will help researchers to further advance the field.

Fatemeh Azimi, Stanislav Frolov, Federico Raue et al.

One-shot Video Object Segmentation~(VOS) is the task of pixel-wise tracking an object of interest within a video sequence, where the segmentation mask of the first frame is given at inference time. In recent years, Recurrent Neural Networks~(RNNs) have been widely used for VOS tasks, but they often suffer from limitations such as drift and error propagation. In this work, we study an RNN-based architecture and address some of these issues by proposing a hybrid sequence-to-sequence architecture named HS2S, utilizing a dual mask propagation strategy that allows incorporating the information obtained from correspondence matching. Our experiments show that augmenting the RNN with correspondence matching is a highly effective solution to reduce the drift problem. The additional information helps the model to predict more accurate masks and makes it robust against error propagation. We evaluate our HS2S model on the DAVIS2017 dataset as well as Youtube-VOS. On the latter, we achieve an improvement of 11.2pp in the overall segmentation accuracy over RNN-based state-of-the-art methods in VOS. We analyze our model's behavior in challenging cases such as occlusion and long sequences and show that our hybrid architecture significantly enhances the segmentation quality in these difficult scenarios.

Fatemeh Azimi, Benjamin Bischke, Sebastian Palacio et al.

Video Object Segmentation (VOS) is an active research area of the visual domain. One of its fundamental sub-tasks is semi-supervised / one-shot learning: given only the segmentation mask for the first frame, the task is to provide pixel-accurate masks for the object over the rest of the sequence. Despite much progress in the last years, we noticed that many of the existing approaches lose objects in longer sequences, especially when the object is small or briefly occluded. In this work, we build upon a sequence-to-sequence approach that employs an encoder-decoder architecture together with a memory module for exploiting the sequential data. We further improve this approach by proposing a model that manipulates multi-scale spatio-temporal information using memory-equipped skip connections. Furthermore, we incorporate an auxiliary task based on distance classification which greatly enhances the quality of edges in segmentation masks. We compare our approach to the state of the art and show considerable improvement in the contour accuracy metric and the overall segmentation accuracy.

Andrey Guzhov, Federico Raue, Jörn Hees et al.

Environmental Sound Classification (ESC) is an active research area in the audio domain and has seen a lot of progress in the past years. However, many of the existing approaches achieve high accuracy by relying on domain-specific features and architectures, making it harder to benefit from advances in other fields (e.g., the image domain). Additionally, some of the past successes have been attributed to a discrepancy of how results are evaluated (i.e., on unofficial splits of the UrbanSound8K (US8K) dataset), distorting the overall progression of the field. The contribution of this paper is twofold. First, we present a model that is inherently compatible with mono and stereo sound inputs. Our model is based on simple log-power Short-Time Fourier Transform (STFT) spectrograms and combines them with several well-known approaches from the image domain (i.e., ResNet, Siamese-like networks and attention). We investigate the influence of cross-domain pre-training, architectural changes, and evaluate our model on standard datasets. We find that our model out-performs all previously known approaches in a fair comparison by achieving accuracies of 97.0 % (ESC-10), 91.5 % (ESC-50) and 84.2 % / 85.4 % (US8K mono / stereo). Second, we provide a comprehensive overview of the actual state of the field, by differentiating several previously reported results on the US8K dataset between official or unofficial splits. For better reproducibility, our code (including any re-implementations) is made available.

Shailza Jolly, Sebastian Palacio, Joachim Folz et al.

In recent years, progress in the Visual Question Answering (VQA) field has largely been driven by public challenges and large datasets. One of the most widely-used of these is the VQA 2.0 dataset, consisting of polar ("yes/no") and non-polar questions. Looking at the question distribution over all answers, we find that the answers "yes" and "no" account for 38 % of the questions, while the remaining 62% are spread over the more than 3000 remaining answers. While several sources of biases have already been investigated in the field, the effects of such an over-representation of polar vs. non-polar questions remain unclear. In this paper, we measure the potential confounding factors when polar and non-polar samples are used jointly to train a baseline VQA classifier, and compare it to an upper bound where the over-representation of polar questions is excluded from the training. Further, we perform cross-over experiments to analyze how well the feature spaces align. Contrary to expectations, we find no evidence of counterproductive effects in the joint training of unbalanced classes. In fact, by exploring the intermediate feature space of visual-text embeddings, we find that the feature space of polar questions already encodes sufficient structure to answer many non-polar questions. Our results indicate that the polar (P) and the non-polar (NP) feature spaces are strongly aligned, hence the expression P $\approx$ NP

Philipp Blandfort, Tushar Karayil, Federico Raue et al.

Growing amounts of online user data motivate the need for automated processing techniques. In case of user ratings, one interesting option is to use neural networks for learning to predict ratings given an item and a user. While training for prediction, such an approach at the same time learns to map each user to a vector, a so-called user embedding. Such embeddings can for example be valuable for estimating user similarity. However, there are various ways how item and user information can be combined in neural networks, and it is unclear how the way of combining affects the resulting embeddings. In this paper, we run an experiment on movie ratings data, where we analyze the effect on embedding quality caused by several fusion strategies in neural networks. For evaluating embedding quality, we propose a novel measure, Pair-Distance Correlation, which quantifies the condition that similar users should have similar embedding vectors. We find that the fusion strategy affects results in terms of both prediction performance and embedding quality. Surprisingly, we find that prediction performance not necessarily reflects embedding quality. This suggests that if embeddings are of interest, the common tendency to select models based on their prediction ability should be reconsidered.

Sebastian Palacio, Joachim Folz, Jörn Hees et al.

We propose a novel way to measure and understand convolutional neural networks by quantifying the amount of input signal they let in. To do this, an autoencoder (AE) was fine-tuned on gradients from a pre-trained classifier with fixed parameters. We compared the reconstructed samples from AEs that were fine-tuned on a set of image classifiers (AlexNet, VGG16, ResNet-50, and Inception~v3) and found substantial differences. The AE learns which aspects of the input space to preserve and which ones to ignore, based on the information encoded in the backpropagated gradients. Measuring the changes in accuracy when the signal of one classifier is used by a second one, a relation of total order emerges. This order depends directly on each classifier's input signal but it does not correlate with classification accuracy or network size. Further evidence of this phenomenon is provided by measuring the normalized mutual information between original images and auto-encoded reconstructions from different fine-tuned AEs. These findings break new ground in the area of neural network understanding, opening a new way to reason, debug, and interpret their results. We present four concrete examples in the literature where observations can now be explained in terms of the input signal that a model uses.

Federico Raue, Andreas Dengel, Thomas M. Breuel et al.

In this paper, we extend a symbolic association framework for being able to handle missing elements in multimodal sequences. The general scope of the work is the symbolic associations of object-word mappings as it happens in language development in infants. In other words, two different representations of the same abstract concepts can associate in both directions. This scenario has been long interested in Artificial Intelligence, Psychology, and Neuroscience. In this work, we extend a recent approach for multimodal sequences (visual and audio) to also cope with missing elements in one or both modalities. Our method uses two parallel Long Short-Term Memories (LSTMs) with a learning rule based on EM-algorithm. It aligns both LSTM outputs via Dynamic Time Warping (DTW). We propose to include an extra step for the combination with the max operation for exploiting the common elements between both sequences. The motivation behind is that the combination acts as a condition selector for choosing the best representation from both LSTMs. We evaluated the proposed extension in the following scenarios: missing elements in one modality (visual or audio) and missing elements in both modalities (visual and sound). The performance of our extension reaches better results than the original model and similar results to individual LSTM trained in each modality.