Sebastian Stober

Ayse Betul Yuce, Sebastian Stober

Electroencephalography (EEG) is a widely used non-invasive technique for measuring brain activity in brain-computer interface (BCI) applications. Supervised EEG decoding models often struggle to generalize across tasks, subjects, and datasets, motivating transformer-based EEG foundation models trained with self-supervised learning. Since transformers are permutation-invariant, they require explicit positional information. Unlike textual tokens, EEG electrodes are spatially distributed across the scalp, raising the question of how electrode positions should be encoded in transformer-based EEG models. In this study, we benchmark five positional encoding strategies within the CBraMod backbone and evaluate them under linear probing and fine-tuning protocols on motor imagery classification and emotion recognition. Our results show that no single strategy consistently outperforms across tasks. Spherical Positional Encoding (SPE) yields strong representations for motor imagery but underperforms on emotion recognition, while Asymmetric Conditional Positional Encoding (ACPE) demonstrates more consistent performance across tasks. These findings suggest that the optimal positional encoding strategy is task-dependent, with no universal solution across EEG decoding scenarios.

Valerie Krug, Raihan Kabir Ratul, Christopher Olson et al.

Machine Learning with Deep Neural Networks (DNNs) has become a successful tool in solving tasks across various fields of application. However, the complexity of DNNs makes it difficult to understand how they solve their learned task. To improve the explainability of DNNs, we adapt methods from neuroscience that analyze complex and opaque systems. Here, we draw inspiration from how neuroscience uses topographic maps to visualize brain activity. To also visualize activations of neurons in DNNs as topographic maps, we research techniques to layout the neurons in a two-dimensional space such that neurons of similar activity are in the vicinity of each other. In this work, we introduce and compare methods to obtain a topographic layout of neurons in a DNN layer. Moreover, we demonstrate how to use topographic activation maps to identify errors or encoded biases and to visualize training processes. Our novel visualization technique improves the transparency of DNN-based decision-making systems and is interpretable without expert knowledge in Machine Learning.

Johann Schmidt, Sebastian Stober

Nearly all state of the art vision models are sensitive to image rotations. Existing methods often compensate for missing inductive biases by using augmented training data to learn pseudo-invariances. Alongside the resource demanding data inflation process, predictions often poorly generalize. The inductive biases inherent to convolutional neural networks allow for translation equivariance through kernels acting parallely to the horizontal and vertical axes of the pixel grid. This inductive bias, however, does not allow for rotation equivariance. We propose a radial beam sampling strategy along with radial kernels operating on these beams to inherently incorporate center-rotation covariance. Together with an angle distance loss, we present a radial beam-based image canonicalization model, short BIC. Our model allows for maximal continuous angle regression and canonicalizes arbitrary center-rotated input images. As a pre-processing model, this enables rotation-invariant vision pipelines with model-agnostic rotation-sensitive downstream predictions. We show that our end-to-end trained angle regressor is able to predict continuous rotation angles on several vision datasets, i.e. FashionMNIST, CIFAR10, COIL100, and LFW.

Arnab Das, Suhita Ghosh, Sebastian Stober

The detailed images produced by Magnetic Resonance Imaging (MRI) provide life-critical information for the diagnosis and treatment of prostate cancer. To provide standardized acquisition, interpretation and usage of the complex MRI images, the PI-RADS v2 guideline was proposed. An automated segmentation following the guideline facilitates consistent and precise lesion detection, staging and treatment. The guideline recommends a division of the prostate into four zones, PZ (peripheral zone), TZ (transition zone), DPU (distal prostatic urethra) and AFS (anterior fibromuscular stroma). Not every zone shares a boundary with the others and is present in every slice. Further, the representations captured by a single model might not suffice for all zones. This motivated us to design a dual-branch convolutional neural network (CNN), where each branch captures the representations of the connected zones separately. Further, the representations from different branches act complementary to each other at the second stage of training, where they are fine-tuned through an unsupervised loss. The loss penalises the difference in predictions from the two branches for the same class. We also incorporate multi-task learning in our framework to further improve the segmentation accuracy. The proposed approach improves the segmentation accuracy of the baseline (mean absolute symmetric distance) by 7.56%, 11.00%, 58.43% and 19.67% for PZ, TZ, DPU and AFS zones respectively.

Suhita Ghosh, Melanie Jouaiti, Arnab Das et al.

Speech anonymisation aims to protect speaker identity by changing personal identifiers in speech while retaining linguistic content. Current methods fail to retain prosody and unique speech patterns found in elderly and pathological speech domains, which is essential for remote health monitoring. To address this gap, we propose a voice conversion-based method (DDSP-QbE) using differentiable digital signal processing and query-by-example. The proposed method, trained with novel losses, aids in disentangling linguistic, prosodic, and domain representations, enabling the model to adapt to uncommon speech patterns. Objective and subjective evaluations show that DDSP-QbE significantly outperforms the voice conversion state-of-the-art concerning intelligibility, prosody, and domain preservation across diverse datasets, pathologies, and speakers while maintaining quality and speaker anonymity. Experts validate domain preservation by analysing twelve clinically pertinent domain attributes.

Johann Schmidt, Sebastian Stober

Fine-grained visual classification (FGVC) remains highly sensitive to geometric variability, where objects appear under arbitrary orientations, scales, and perspective distortions. While equivariant architectures address this issue, they typically require substantial computational resources and restrict the hypothesis space. We revisit Spatial Transformer Networks (STNs) as a canonicalization tool for transformer-based vision pipelines, emphasizing their flexibility, backbone-agnostic nature, and lack of architectural constraints. We propose a probabilistic, component-wise extension that improves robustness. Specifically, we decompose affine transformations into rotation, scaling, and shearing, and regress each component under geometric constraints using a shared localization encoder. To capture uncertainty, we model each component with a Gaussian variational posterior and perform sampling-based canonicalization during inference.A novel component-wise alignment loss leverages augmentation parameters to guide spatial alignment. Experiments on challenging moth classification benchmarks demonstrate that our method consistently improves robustness compared to other STNs.

Suhita Ghosh, Yamini Sinha, Sebastian Stober

Differentiable Digital Signal Processing (DDSP) pipelines for voice conversion rely on subtractive synthesis, where a periodic excitation signal is shaped by a learned spectral envelope to reconstruct the target voice. In DDSP-QbE, the excitation is generated via phase accumulation, producing a sawtooth-like waveform whose abrupt discontinuities introduce aliasing artefacts that manifest perceptually as buzziness and spectral distortion, particularly at higher fundamental frequencies. We propose two targeted improvements to the excitation stage of the DDSP-QbE subtractive synthesizer. First, we incorporate explicit voicing detection to gate the harmonic excitation, suppressing the periodic component in unvoiced regions and replacing it with filtered noise, thereby avoiding aliased harmonic content where it is most perceptually disruptive. Second, we apply Polynomial Band-Limited Step (PolyBLEP) correction to the phase-accumulated oscillator, substituting the hard waveform discontinuity at each phase wrap with a smooth polynomial residual that cancels alias-generating components without oversampling or spectral truncation. Together, these modifications yield a cleaner harmonic roll-off, reduced high-frequency artefacts, and improved perceptual naturalness, as measured by MOS. The proposed approach is lightweight, differentiable, and integrates seamlessly into the existing DDSP-QbE training pipeline with no additional learnable parameters.

Dominik Eckert, Christopher Syben, Christian Hümmer et al.

The progression of X-ray technology introduces diverse image styles that need to be adapted to the preferences of radiologists. To support this task, we introduce a novel deep learning-based metric that quantifies style differences of non-matching image pairs. At the heart of our metric is an encoder capable of generating X-ray image style representations. This encoder is trained without any explicit knowledge of style distances by exploiting Simple Siamese learning. During inference, the style representations produced by the encoder are used to calculate a distance metric for non-matching image pairs. Our experiments investigate the proposed concept for a disclosed reproducible and a proprietary image processing pipeline along two dimensions: First, we use a t-distributed stochastic neighbor embedding (t-SNE) analysis to illustrate that the encoder outputs provide meaningful and discriminative style representations. Second, the proposed metric calculated from the encoder outputs is shown to quantify style distances for non-matching pairs in good alignment with the human perception. These results confirm that our proposed method is a promising technique to quantify style differences, which can be used for guided style selection as well as automatic optimization of image pipeline parameters.

Tim W. Dornis, Sebastian Stober

Training generative AI models requires extensive amounts of data. A common practice is to collect such data through web scraping. Yet, much of what has been and is collected is copyright protected. Its use may be copyright infringement. In the USA, AI developers rely on "fair use" and in Europe, the prevailing view is that the exception for "Text and Data Mining" (TDM) applies. In a recent interdisciplinary tandem-study, we have argued in detail that this is actually not the case because generative AI training fundamentally differs from TDM. In this article, we share our main findings and the implications for both public and corporate research on generative models. We further discuss how the phenomenon of training data memorization leads to copyright issues independently from the "fair use" and TDM exceptions. Finally, we outline how the ISMIR could contribute to the ongoing discussion about fair practices with respect to generative AI that satisfy all stakeholders.

Dominik Eckert, Ludwig Ritschl, Christopher Syben et al.

Radiologists have preferred visual impressions or 'styles' of X-ray images that are manually adjusted to their needs to support their diagnostic performance. In this work, we propose an automatic and interpretable X-ray style transfer by introducing a trainable version of the Local Laplacian Filter (LLF). From the shape of the LLF's optimized remap function, the characteristics of the style transfer can be inferred and reliability of the algorithm can be ensured. Moreover, we enable the LLF to capture complex X-ray style features by replacing the remap function with a Multi-Layer Perceptron (MLP) and adding a trainable normalization layer. We demonstrate the effectiveness of the proposed method by transforming unprocessed mammographic X-ray images into images that match the style of target mammograms and achieve a Structural Similarity Index (SSIM) of 0.94 compared to 0.82 of the baseline LLF style transfer method from Aubry et al.

Maral Ebrahimzadeh, Gilberto Bernardes, Sebastian Stober

State-of-the-art symbolic music generation models have recently achieved remarkable output quality, yet explicit control over compositional features, such as tonal tension, remains challenging. We propose a novel approach that integrates a computational tonal tension model, based on tonal interval vector analysis, into a Transformer framework. Our method employs a two-level beam search strategy during inference. At the token level, generated candidates are re-ranked using model probability and diversity metrics to maintain overall quality. At the bar level, a tension-based re-ranking is applied to ensure that the generated music aligns with a desired tension curve. Objective evaluations indicate that our approach effectively modulates tonal tension, and subjective listening tests confirm that the system produces outputs that align with the target tension. These results demonstrate that explicit tension conditioning through a dual-level beam search provides a powerful and intuitive tool to guide AI-generated music. Furthermore, our experiments demonstrate that our method can generate multiple distinct musical interpretations under the same tension condition.

Johann Schmidt, Sebastian Stober

Fine-grained visual classification (FGVC) tasks, such as insect and bird identification, demand sensitivity to subtle visual cues while remaining robust to spatial transformations. A key challenge is handling geometric biases and noise, such as different orientations and scales of objects. Existing remedies rely on heavy data augmentation, which demands powerful models, or on equivariant architectures, which constrain expressivity and add cost. Canonicalization offers an alternative by shielding such biases from the downstream model. In practice, such functions are often obtained using canonicalization priors, which assume aligned training data. Unfortunately, real-world datasets never fulfill this assumption, causing the obtained canonicalizer to be brittle. We propose a bootstrapping algorithm that iteratively re-aligns training samples by progressively reducing variance and recovering the alignment assumption. We establish convergence guarantees under mild conditions for arbitrary compact groups, and show on four FGVC benchmarks that our method consistently outperforms equivariant, and canonicalization baselines while performing on par with augmentation.

Johann Schmidt, Sebastian Stober, Joachim Denzler et al.

Fine-grained visual classification (FGVC) requires distinguishing between visually similar categories through subtle, localized features - a task that remains challenging due to high intra-class variability and limited inter-class differences. Existing part-based methods often rely on complex localization networks that learn mappings from pixel to sample space, requiring a deep understanding of image content while limiting feature utility for downstream tasks. In addition, sampled points frequently suffer from high spatial redundancy, making it difficult to quantify the optimal number of required parts. Inspired by human saccadic vision, we propose a two-stage process that first extracts peripheral features (coarse view) and generates a sample map, from which fixation patches are sampled and encoded in parallel using a weight-shared encoder. We employ contextualized selective attention to weigh the impact of each fixation patch before fusing peripheral and focus representations. To prevent spatial collapse - a common issue in part-based methods - we utilize non-maximum suppression during fixation sampling to eliminate redundancy. Comprehensive evaluation on standard FGVC benchmarks (CUB-200-2011, NABirds, Food-101 and Stanford-Dogs) and challenging insect datasets (EU-Moths, Ecuador-Moths and AMI-Moths) demonstrates that our method achieves comparable performance to state-of-the-art approaches while consistently outperforming our baseline encoder.

Benjamin El-Zein, Dominik Eckert, Andreas Fieselmann et al.

Collimation in X-ray imaging restricts exposure to the region-of-interest (ROI) and minimizes the radiation dose applied to the patient. The detection of collimator shadows is an essential image-based preprocessing step in digital radiography posing a challenge when edges get obscured by scattered X-ray radiation. Regardless, the prior knowledge that collimation forms polygonal-shaped shadows is evident. For this reason, we introduce a deep learning-based segmentation that is inherently constrained to its geometry. We achieve this by incorporating a differentiable Hough transform-based network to detect the collimation borders and enhance its capability to extract the information about the ROI center. During inference, we combine the information of both tasks to enable the generation of refined, line-constrained segmentation masks. We demonstrate robust reconstruction of collimated regions achieving median Hausdorff distances of 4.3-5.0mm on diverse test sets of real Xray images. While this application involves at most four shadow borders, our method is not fundamentally limited by a specific number of edges.

Johannes Schleiss, Anke Manukjan, Michelle Ines Bieber et al.

As Artificial Intelligence (AI) increasingly impacts professional practice, there is a growing need to AI-related competencies into higher education curricula. However, research on the implementation of AI education within study programs remains limited and requires new forms of collaboration across disciplines. This study addresses this gap and explores perspectives on interdisciplinary curriculum development through the lens of different stakeholders. In particular, we examine the case of curriculum development for a novel undergraduate program in AI in engineering. The research uses a mixed methods approach, combining quantitative curriculum mapping with qualitative focus group interviews. In addition to assessing the alignment of the curriculum with the targeted competencies, the study also examines the perceived quality, consistency, practicality and effectiveness from both academic and industry perspectives, as well as differences in perceptions between educators who were involved in the development and those who were not. The findings provide a practical understanding of the outcomes of interdisciplinary AI curriculum development and contribute to a broader understanding of how educator participation in curriculum development influences perceptions of quality aspects. It also advances the field of AI education by providing a reference point and insights for further interdisciplinary curriculum developments in response to evolving industry needs.

Suhita Ghosh, Melanie Jouaiti, Jan-Ole Perschewski et al.

Detecting and segmenting dysfluencies is crucial for effective speech therapy and real-time feedback. However, most methods only classify dysfluencies at the utterance level. We introduce StutterCut, a semi-supervised framework that formulates dysfluency segmentation as a graph partitioning problem, where speech embeddings from overlapping windows are represented as graph nodes. We refine the connections between nodes using a pseudo-oracle classifier trained on weak (utterance-level) labels, with its influence controlled by an uncertainty measure from Monte Carlo dropout. Additionally, we extend the weakly labelled FluencyBank dataset by incorporating frame-level dysfluency boundaries for four dysfluency types. This provides a more realistic benchmark compared to synthetic datasets. Experiments on real and synthetic datasets show that StutterCut outperforms existing methods, achieving higher F1 scores and more precise stuttering onset detection.

Valerie Krug, Sebastian Stober

Deep Learning models have achieved remarkable success. Training them is often accelerated by building on top of pre-trained models which poses the risk of perpetuating encoded biases. Here, we investigate biases in the representations of commonly used ImageNet classifiers for facial images while considering intersections of sensitive variables age, race and gender. To assess the biases, we use linear classifier probes and visualize activations as topographic maps. We find that representations in ImageNet classifiers particularly allow differentiation between ages. Less strongly pronounced, the models appear to associate certain ethnicities and distinguish genders in middle-aged groups.

Johann Schmidt, Frank Dreyer, Sayed Abid Hashimi et al.

Traffic signal control plays a crucial role in urban mobility. However, existing methods often struggle to generalize beyond their training environments to unseen scenarios with varying traffic dynamics. We present TransferLight, a novel framework designed for robust generalization across road-networks, diverse traffic conditions and intersection geometries. At its core, we propose a log-distance reward function, offering spatially-aware signal prioritization while remaining adaptable to varied lane configurations - overcoming the limitations of traditional pressure-based rewards. Our hierarchical, heterogeneous, and directed graph neural network architecture effectively captures granular traffic dynamics, enabling transferability to arbitrary intersection layouts. Using a decentralized multi-agent approach, global rewards, and novel state transition priors, we develop a single, weight-tied policy that scales zero-shot to any road network without re-training. Through domain randomization during training, we additionally enhance generalization capabilities. Experimental results validate TransferLight's superior performance in unseen scenarios, advancing practical, generalizable intelligent transportation systems to meet evolving urban traffic demands.

Suhita Ghosh, Tim Thiele, Frederic Lorbeer et al.

The increasing use of cloud-based speech assistants has heightened the need for effective speech anonymization, which aims to obscure a speaker's identity while retaining critical information for subsequent tasks. One approach to achieving this is through voice conversion. While existing methods often emphasize complex architectures and training techniques, our research underscores the importance of loss functions inspired by the human auditory system. Our proposed loss functions are model-agnostic, incorporating handcrafted and deep learning-based features to effectively capture quality representations. Through objective and subjective evaluations, we demonstrate that a VQVAE-based model, enhanced with our perception-driven losses, surpasses the vanilla model in terms of naturalness, intelligibility, and prosody while maintaining speaker anonymity. These improvements are consistently observed across various datasets, languages, target speakers, and genders.

Johann Schmidt, Sebastian Stober

Deep neural networks are applied in more and more areas of everyday life. However, they still lack essential abilities, such as robustly dealing with spatially transformed input signals. Approaches to mitigate this severe robustness issue are limited to two pathways: Either models are implicitly regularised by increased sample variability (data augmentation) or explicitly constrained by hard-coded inductive biases. The limiting factor of the former is the size of the data space, which renders sufficient sample coverage intractable. The latter is limited by the engineering effort required to develop such inductive biases for every possible scenario. Instead, we take inspiration from human behaviour, where percepts are modified by mental or physical actions during inference. We propose a novel technique to emulate such an inference process for neural nets. This is achieved by traversing a sparsified inverse transformation tree during inference using parallel energy-based evaluations. Our proposed inference algorithm, called Inverse Transformation Search (ITS), is model-agnostic and equips the model with zero-shot pseudo-invariance to spatially transformed inputs. We evaluated our method on several benchmark datasets, including a synthesised ImageNet test set. ITS outperforms the utilised baselines on all zero-shot test scenarios.

André Ofner, Sebastian Stober

This paper deals with differentiable dynamical models congruent with neural process theories that cast brain function as the hierarchical refinement of an internal generative model explaining observations. Our work extends existing implementations of gradient-based predictive coding with automatic differentiation and allows to integrate deep neural networks for non-linear state parameterization. Gradient-based predictive coding optimises inferred states and weights locally in for each layer by optimising precision-weighted prediction errors that propagate from stimuli towards latent states. Predictions flow backwards, from latent states towards lower layers. The model suggested here optimises hierarchical and dynamical predictions of latent states. Hierarchical predictions encode expected content and hierarchical structure. Dynamical predictions capture changes in the encoded content along with higher order derivatives. Hierarchical and dynamical predictions interact and address different aspects of the same latent states. We apply the model to various perception and planning tasks on sequential data and show their mutual dependence. In particular, we demonstrate how learning sampling distances in parallel address meaningful locations data sampled at discrete time steps. We discuss possibilities to relax the assumption of linear hierarchies in favor of more flexible graph structure with emergent properties. We compare the granular structure of the model with canonical microcircuits describing predictive coding in biological networks and review the connection to Markov Blankets as a tool to characterize modularity. A final section sketches out ideas for efficient perception and planning in nested spatio-temporal hierarchies.

André Ofner, Sebastian Stober

We present PredProp, a method for optimization of weights and states in predictive coding networks (PCNs) based on the precision of propagated errors and neural activity. PredProp jointly addresses inference and learning via stochastic gradient descent and adaptively weights parameter updates by approximate curvature. Due to the relation between propagated error covariance and the Fisher information matrix, PredProp implements approximate Natural Gradient Descent. We demonstrate PredProp's effectiveness in the context of dense decoder networks and simple image benchmark datasets. We found that PredProp performs favorably over Adam, a widely used adaptive learning rate optimizer in the tested configurations. Furthermore, available optimization methods for weight parameters benefit from using PredProp's error precision during inference. Since hierarchical predictive coding layers are optimised individually using local errors, the required precisions factorize over hierarchical layers. Extending beyond classical PCNs with a single set of decoder layers per hierarchical layer, we also generalize PredProp to deep neural networks in each PCN layer by additionally factorizing over the weights in each PCN layer.

Andre Ofner, Raihan Kabir Ratul, Suhita Ghosh et al.

There is an increasing convergence between biologically plausible computational models of inference and learning with local update rules and the global gradient-based optimization of neural network models employed in machine learning. One particularly exciting connection is the correspondence between the locally informed optimization in predictive coding networks and the error backpropagation algorithm that is used to train state-of-the-art deep artificial neural networks. Here we focus on the related, but still largely under-explored connection between precision weighting in predictive coding networks and the Natural Gradient Descent algorithm for deep neural networks. Precision-weighted predictive coding is an interesting candidate for scaling up uncertainty-aware optimization -- particularly for models with large parameter spaces -- due to its distributed nature of the optimization process and the underlying local approximation of the Fisher information metric, the adaptive learning rate that is central to Natural Gradient Descent. Here, we show that hierarchical predictive coding networks with learnable precision indeed are able to solve various supervised and unsupervised learning tasks with performance comparable to global backpropagation with natural gradients and outperform their classical gradient descent counterpart on tasks where high amounts of noise are embedded in data or label inputs. When applied to unsupervised auto-encoding of image inputs, the deterministic network produces hierarchically organized and disentangled embeddings, hinting at the close connections between predictive coding and hierarchical variational inference.

Indrani Sarkar, Indranil Maji, Charitha Omprakash et al.

The assessment of laboratory animal behavior is of central interest in modern neuroscience research. Behavior is typically studied in terms of pose changes, which are ideally captured in three dimensions. This requires triangulation over a multi-camera system which view the animal from different angles. However, this is challenging in realistic laboratory setups due to occlusions and other technical constrains. Here we propose the usage of lift-pose models that allow for robust 3D pose estimation of freely moving rodents from a single view camera view. To obtain high-quality training data for the pose-lifting, we first perform geometric calibration in a camera setup involving bottom as well as side views of the behaving animal. We then evaluate the performance of two previously proposed model architectures under given inference perspectives and conclude that reliable 3D pose inference can be obtained using temporal convolutions. With this work we would like to contribute to a more robust and diverse behavior tracking of freely moving rodents for a wide range of experiments and setups in the neuroscience community.

Anneke Meyer, Suhita Ghosh, Daniel Schindele et al.

Various convolutional neural network (CNN) based concepts have been introduced for the prostate's automatic segmentation and its coarse subdivision into transition zone (TZ) and peripheral zone (PZ). However, when targeting a fine-grained segmentation of TZ, PZ, distal prostatic urethra (DPU) and the anterior fibromuscular stroma (AFS), the task becomes more challenging and has not yet been solved at the level of human performance. One reason might be the insufficient amount of labeled data for supervised training. Therefore, we propose to apply a semi-supervised learning (SSL) technique named uncertainty-aware temporal self-learning (UATS) to overcome the expensive and time-consuming manual ground truth labeling. We combine the SSL techniques temporal ensembling and uncertainty-guided self-learning to benefit from unlabeled images, which are often readily available. Our method significantly outperforms the supervised baseline and obtained a Dice coefficient (DC) of up to 78.9% , 87.3%, 75.3%, 50.6% for TZ, PZ, DPU and AFS, respectively. The obtained results are in the range of human inter-rater performance for all structures. Moreover, we investigate the method's robustness against noise and demonstrate the generalization capability for varying ratios of labeled data and on other challenging tasks, namely the hippocampus and skin lesion segmentation. UATS achieved superiority segmentation quality compared to the supervised baseline, particularly for minimal amounts of labeled data.

Soumick Chatterjee, Fatima Saad, Chompunuch Sarasaen et al.

The outbreak of COVID-19 has shocked the entire world with its fairly rapid spread and has challenged different sectors. One of the most effective ways to limit its spread is the early and accurate diagnosing infected patients. Medical imaging, such as X-ray and Computed Tomography (CT), combined with the potential of Artificial Intelligence (AI), plays an essential role in supporting medical personnel in the diagnosis process. Thus, in this article five different deep learning models (ResNet18, ResNet34, InceptionV3, InceptionResNetV2 and DenseNet161) and their ensemble, using majority voting have been used to classify COVID-19, pneumoniæ and healthy subjects using chest X-ray images. Multilabel classification was performed to predict multiple pathologies for each patient, if present. Firstly, the interpretability of each of the networks was thoroughly studied using local interpretability methods - occlusion, saliency, input X gradient, guided backpropagation, integrated gradients, and DeepLIFT, and using a global technique - neuron activation profiles. The mean Micro-F1 score of the models for COVID-19 classifications ranges from 0.66 to 0.875, and is 0.89 for the ensemble of the network models. The qualitative results showed that the ResNets were the most interpretable models. This research demonstrates the importance of using interpretability methods to compare different models before making a decision regarding the best performing model.

Andreas Krug, Sebastian Stober

Deep Learning based Automatic Speech Recognition (ASR) models are very successful, but hard to interpret. To gain better understanding of how Artificial Neural Networks (ANNs) accomplish their tasks, introspection methods have been proposed. Adapting such techniques from computer vision to speech recognition is not straight-forward, because speech data is more complex and less interpretable than image data. In this work, we introduce Gradient-adjusted Neuron Activation Profiles (GradNAPs) as means to interpret features and representations in Deep Neural Networks. GradNAPs are characteristic responses of ANNs to particular groups of inputs, which incorporate the relevance of neurons for prediction. We show how to utilize GradNAPs to gain insight about how data is processed in ANNs. This includes different ways of visualizing features and clustering of GradNAPs to compare embeddings of different groups of inputs in any layer of a given network. We demonstrate our proposed techniques using a fully-convolutional ASR model.

Andreas Krug, Sebastian Stober

The uninformative ordering of artificial neurons in Deep Neural Networks complicates visualizing activations in deeper layers. This is one reason why the internal structure of such models is very unintuitive. In neuroscience, activity of real brains can be visualized by highlighting active regions. Inspired by those techniques, we train a convolutional speech recognition model, where filters are arranged in a 2D grid and neighboring filters are similar to each other. We show, how those topographic filter maps visualize artificial neuron activations more intuitively. Moreover, we investigate, whether this causes phoneme-responsive neurons to be grouped in certain regions of the topographic map.

Roshan Rane, Edit Szügyi, Vageesh Saxena et al.

PredNet, a deep predictive coding network developed by Lotter et al., combines a biologically inspired architecture based on the propagation of prediction error with self-supervised representation learning in video. While the architecture has drawn a lot of attention and various extensions of the model exist, there is a lack of a critical analysis. We fill in the gap by evaluating PredNet both as an implementation of the predictive coding theory and as a self-supervised video prediction model using a challenging video action classification dataset. We design an extended model to test if conditioning future frame predictions on the action class of the video improves the model performance. We show that PredNet does not yet completely follow the principles of predictive coding. The proposed top-down conditioning leads to a performance gain on synthetic data, but does not scale up to the more complex real-world action classification dataset. Our analysis is aimed at guiding future research on similar architectures based on the predictive coding theory.

André Ofner, Sebastian Stober

We describe a framework of hybrid cognition by formulating a hybrid cognitive agent that performs hierarchical active inference across a human and a machine part. We suggest that, in addition to enhancing human cognitive functions with an intelligent and adaptive interface, integrated cognitive processing could accelerate emergent properties within artificial intelligence. To establish this, a machine learning part learns to integrate into human cognition by explaining away multi-modal sensory measurements from the environment and physiology simultaneously with the brain signal. With ongoing training, the amount of predictable brain signal increases. This lends the agent the ability to self-supervise on increasingly high levels of cognitive processing in order to further minimize surprise in predicting the brain signal. Furthermore, with increasing level of integration, the access to sensory information about environment and physiology is substituted with access to their representation in the brain. While integrating into a joint embodiment of human and machine, human action and perception are treated as the machine's own. The framework can be implemented with invasive as well as non-invasive sensors for environment, body and brain interfacing. Online and offline training with different machine learning approaches are thinkable. Building on previous research on shared representation learning, we suggest a first implementation leading towards hybrid active inference with non-invasive brain interfacing and state of the art probabilistic deep learning methods. We further discuss how implementation might have effect on the meta-cognitive abilities of the described agent and suggest that with adequate implementation the machine part can continue to execute and build upon the learned cognitive processes autonomously.

Julius Kunze, Louis Kirsch, Ilia Kurenkov et al.

End-to-end training of automated speech recognition (ASR) systems requires massive data and compute resources. We explore transfer learning based on model adaptation as an approach for training ASR models under constrained GPU memory, throughput and training data. We conduct several systematic experiments adapting a Wav2Letter convolutional neural network originally trained for English ASR to the German language. We show that this technique allows faster training on consumer-grade resources while requiring less training data in order to achieve the same accuracy, thereby lowering the cost of training ASR models in other languages. Model introspection revealed that small adaptations to the network's weights were sufficient for good performance, especially for inner layers.

Sebastian Stober, Avital Sternin, Adrian M. Owen et al.

We introduce and compare several strategies for learning discriminative features from electroencephalography (EEG) recordings using deep learning techniques. EEG data are generally only available in small quantities, they are high-dimensional with a poor signal-to-noise ratio, and there is considerable variability between individual subjects and recording sessions. Our proposed techniques specifically address these challenges for feature learning. Cross-trial encoding forces auto-encoders to focus on features that are stable across trials. Similarity-constraint encoders learn features that allow to distinguish between classes by demanding that two trials from the same class are more similar to each other than to trials from other classes. This tuple-based training approach is especially suitable for small datasets. Hydra-nets allow for separate processing pathways adapting to subsets of a dataset and thus combine the advantages of individual feature learning (better adaptation of early, low-level processing) with group model training (better generalization of higher-level processing in deeper layers). This way, models can, for instance, adapt to each subject individually to compensate for differences in spatial patterns due to anatomical differences or variance in electrode positions. The different techniques are evaluated using the publicly available OpenMIIR dataset of EEG recordings taken while participants listened to and imagined music.