Nils Körber

Yasser Taha, Grégoire Montavon, Nils Körber

While machine learning (ML) architectures have evolved rapidly to account for complex data, loss functions like cross-entropy remain mostly structure-agnostic in many real-world applications. However, the `class-symmetric' nature of these standard losses fundamentally limits the ability of ML models to exploit structural relationships between classes, particularly when facing structured noise. We propose \textsc{Conveyance}, a new classification approach and associated loss function tailored to structured class spaces. It allows users to encode graph-like relations between classes without having to define complex joint distributions or manually tune utility matrices.Technically, our loss function operates by maximizing two separate margins over distinct class partitions, while preserving formal properties such as monotonicity and partial convexity. We demonstrate the versatility and effectiveness of our method by applying it to hierarchical classification, ordinal regression, and multiple instance learning. Across these tasks, \textsc{Conveyance} either matches or exceeds the performance of specialized baselines, thereby offering a unified solution for structured class spaces.

Yasser Taha, Grégoire Montavon, Nils Körber

Modern deep learning faces significant challenges with noisy labels, class ambiguity, as well as the need to robustly reject out-of-distribution or corrupted samples. In this work, we propose a unified framework based on the concept of a "drainage node'' which we add at the output of the network. The node serves to reallocate probability mass toward uncertainty, while preserving desirable properties such as end-to-end training and differentiability. This mechanism provides a natural escape route for highly ambiguous, anomalous, or noisy samples, particularly relevant for instance-dependent and asymmetric label noise. In systematic experiments involving the addition of varying proportions of instance-dependent noise or asymmetric noise to CIFAR-10/100 labels, our drainage formulation achieves an accuracy increase of up to 9\% over existing approaches in the high-noise regime. Our results on real-world datasets, such as mini-WebVision, mini-ImageNet and Clothing-1M, match or surpass existing state-of-the-art methods. Qualitative analysis reveals a denoising effect, where the drainage neuron consistently absorbs corrupt, mislabeled, or outlier data, leading to more stable decision boundaries. Furthermore, our drainage formulation enables applications well beyond classification, with immediate benefits for web-scale, semi-supervised dataset cleaning, and open-set applications.

Nils Körber

Visual perception is driven by the focus on relevant aspects in the surrounding world. To transfer this observation to the digital information processing of computers, attention mechanisms have been introduced to highlight salient image regions. Here, we introduce a parameter-free attention mechanism called PfAAM, that is a simple yet effective module. It can be plugged into various convolutional neural network architectures with a little computational overhead and without affecting model size. PfAAM was tested on multiple architectures for classification and segmentic segmentation leading to improved model performance for all tested cases. This demonstrates its wide applicability as a general easy-to-use module for computer vision tasks. The implementation of PfAAM can be found on https://github.com/nkoerb/pfaam.

Jean-Baptiste Escudié, Benjamin Barnes, Stefan Meisegeier et al.

There is no consensus in the field of synthetic data on concise metrics for quality evaluations or benchmarks on large health datasets, such as historical epidemiological data. This study presents an evaluation of seven recent models from major machine learning families. The models were evaluated using four different datasets, each with a distinct scale. To ensure a fair comparison, we systematically tuned the hyperparameters of each model for each dataset. We propose a methodology for evaluating the fidelity of synthesized joint distributions, aligning metrics with visualization on a single plot. This method is applicable to any dataset and is complemented by a domain-specific analysis of the German Cancer Registries' epidemiological dataset. The analysis reveals the challenges models face in strictly adhering to the medical domain. We hope this approach will serve as a foundational framework for guiding the selection of synthesizers and remain accessible to all stakeholders involved in releasing synthetic datasets.

Nils Körber, Silvan Wehrli, Christopher Irrgang

With the release of ChatGPT and other large language models (LLMs) the discussion about the intelligence, possibilities, and risks, of current and future models have seen large attention. This discussion included much debated scenarios about the imminent rise of so-called "super-human" AI, i.e., AI systems that are orders of magnitude smarter than humans. In the spirit of Alan Turing, there is no doubt that current state-of-the-art language models already pass his famous test. Moreover, current models outperform humans in several benchmark tests, so that publicly available LLMs have already become versatile companions that connect everyday life, industry and science. Despite their impressive capabilities, LLMs sometimes fail completely at tasks that are thought to be trivial for humans. In other cases, the trustworthiness of LLMs becomes much more elusive and difficult to evaluate. Taking the example of academia, language models are capable of writing convincing research articles on a given topic with only little input. Yet, the lack of trustworthiness in terms of factual consistency or the existence of persistent hallucinations in AI-generated text bodies has led to a range of restrictions for AI-based content in many scientific journals. In view of these observations, the question arises as to whether the same metrics that apply to human intelligence can also be applied to computational methods and has been discussed extensively. In fact, the choice of metrics has already been shown to dramatically influence assessments on potential intelligence emergence. Here, we argue that the intelligence of LLMs should not only be assessed by task-specific statistical metrics, but separately in terms of qualitative and quantitative measures.

Shakhnaz Akhmedova, Nils Körber

Neural networks are trained by minimizing a loss function that defines the discrepancy between the predicted model output and the target value. The selection of the loss function is crucial to achieve task-specific behaviour and highly influences the capability of the model. A variety of loss functions have been proposed for a wide range of tasks affecting training and model performance. For classification tasks, the cross entropy is the de-facto standard and usually the first choice. Here, we try to experimentally challenge the well-known loss functions, including cross entropy (CE) loss, by utilizing the genetic programming (GP) approach, a population-based evolutionary algorithm. GP constructs loss functions from a set of operators and leaf nodes and these functions are repeatedly recombined and mutated to find an optimal structure. Experiments were carried out on different small-sized datasets CIFAR-10, CIFAR-100 and Fashion-MNIST using an Inception model. The 5 best functions found were evaluated for different model architectures on a set of standard datasets ranging from 2 to 102 classes and very different sizes. One function, denoted as Next Generation Loss (NGL), clearly stood out showing same or better performance for all tested datasets compared to CE. To evaluate the NGL function on a large-scale dataset, we tested its performance on the Imagenet-1k dataset where it showed improved top-1 accuracy compared to models trained with identical settings and other losses. Finally, the NGL was trained on a segmentation downstream task for Pascal VOC 2012 and COCO-Stuff164k datasets improving the underlying model performance.

Shakhnaz Akhmedova, Nils Körber

Generative adversarial networks (GANs) are machine learning models that are used to estimate the underlying statistical structure of a given dataset and as a result can be used for a variety of tasks such as image generation or anomaly detection. Despite their initial simplicity, designing an effective loss function for training GANs remains challenging, and various loss functions have been proposed aiming to improve the performance and stability of the generative models. In this study, loss function design for GANs is presented as an optimization problem solved using the genetic programming (GP) approach. Initial experiments were carried out using small Deep Convolutional GAN (DCGAN) model and the MNIST dataset, in order to search experimentally for an improved loss function. The functions found were evaluated on CIFAR10, with the best function, named GANetic loss, showing exceptionally better performance and stability compared to the losses commonly used for GAN training. To further evalute its general applicability on more challenging problems, GANetic loss was applied for two medical applications: image generation and anomaly detection. Experiments were performed with histopathological, gastrointestinal or glaucoma images to evaluate the GANetic loss in medical image generation, resulting in improved image quality compared to the baseline models. The GANetic Loss used for polyp and glaucoma images showed a strong improvement in the detection of anomalies. In summary, the GANetic loss function was evaluated on multiple datasets and applications where it consistently outperforms alternative loss functions. Moreover, GANetic loss leads to stable training and reproducible results, a known weak spot of GANs.