Christian Igel

Abdelrahman Mohamed, Hung-yi Lee, Lasse Borgholt et al. · meta-ai, mit

Although supervised deep learning has revolutionized speech and audio processing, it has necessitated the building of specialist models for individual tasks and application scenarios. It is likewise difficult to apply this to dialects and languages for which only limited labeled data is available. Self-supervised representation learning methods promise a single universal model that would benefit a wide variety of tasks and domains. Such methods have shown success in natural language processing and computer vision domains, achieving new levels of performance while reducing the number of labels required for many downstream scenarios. Speech representation learning is experiencing similar progress in three main categories: generative, contrastive, and predictive methods. Other approaches rely on multi-modal data for pre-training, mixing text or visual data streams with speech. Although self-supervised speech representation is still a nascent research area, it is closely related to acoustic word embedding and learning with zero lexical resources, both of which have seen active research for many years. This review presents approaches for self-supervised speech representation learning and their connection to other research areas. Since many current methods focus solely on automatic speech recognition as a downstream task, we review recent efforts on benchmarking learned representations to extend the application beyond speech recognition.

Matthias Freiberger, Peter Kun, Christian Igel et al.

Models leveraging both visual and textual data such as Contrastive Language-Image Pre-training (CLIP), are the backbone of many recent advances in artificial intelligence. In this work, we show that despite their versatility, such models are vulnerable to what we refer to as fooling master images. Fooling master images are capable of maximizing the confidence score of a CLIP model for a significant number of widely varying prompts, while being either unrecognizable or unrelated to the attacked prompts for humans. The existence of such images is problematic as it could be used by bad actors to maliciously interfere with CLIP-trained image retrieval models in production with comparably small effort as a single image can attack many different prompts. We demonstrate how fooling master images for CLIP (CLIPMasterPrints) can be mined using stochastic gradient descent, projected gradient descent, or blackbox optimization. Contrary to many common adversarial attacks, the blackbox optimization approach allows us to mine CLIPMasterPrints even when the weights of the model are not accessible. We investigate the properties of the mined images, and find that images trained on a small number of image captions generalize to a much larger number of semantically related captions. We evaluate possible mitigation strategies, where we increase the robustness of the model and introduce an approach to automatically detect CLIPMasterPrints to sanitize the input of vulnerable models. Finally, we find that vulnerability to CLIPMasterPrints is related to a modality gap in contrastive pre-trained multi-modal networks. Code available at https://github.com/matfrei/CLIPMasterPrints.

Pedram Bakhtiarifard, Christian Igel, Raghavendra Selvan

Energy consumption from the selection, training, and deployment of deep learning models has seen a significant uptick recently. This work aims to facilitate the design of energy-efficient deep learning models that require less computational resources and prioritize environmental sustainability by focusing on the energy consumption. Neural architecture search (NAS) benefits from tabular benchmarks, which evaluate NAS strategies cost-effectively through precomputed performance statistics. We advocate for including energy efficiency as an additional performance criterion in NAS. To this end, we introduce an enhanced tabular benchmark encompassing data on energy consumption for varied architectures. The benchmark, designated as EC-NAS, has been made available in an open-source format to advance research in energy-conscious NAS. EC-NAS incorporates a surrogate model to predict energy consumption, aiding in diminishing the energy expenditure of the dataset creation. Our findings emphasize the potential of EC-NAS by leveraging multi-objective optimization algorithms, revealing a balance between energy usage and accuracy. This suggests the feasibility of identifying energy-lean architectures with little or no compromise in performance.

Dustin Wright, Christian Igel, Gabrielle Samuel et al.

Artificial intelligence (AI) is currently spearheaded by machine learning (ML) methods such as deep learning which have accelerated progress on many tasks thought to be out of reach of AI. These recent ML methods are often compute hungry, energy intensive, and result in significant green house gas emissions, a known driver of anthropogenic climate change. Additionally, the platforms on which ML systems run are associated with environmental impacts that go beyond the energy consumption driven carbon emissions. The primary solution lionized by both industry and the ML community to improve the environmental sustainability of ML is to increase the compute and energy efficiency with which ML systems operate. In this perspective, we argue that it is time to look beyond efficiency in order to make ML more environmentally sustainable. We present three high-level discrepancies between the many variables that influence the efficiency of ML and the environmental sustainability of ML. Firstly, we discuss how compute efficiency does not imply energy efficiency or carbon efficiency. Second, we present the unexpected effects of efficiency on operational emissions throughout the ML model life cycle. And, finally, we explore the broader environmental impacts that are not accounted by efficiency. These discrepancies show as to why efficiency alone is not enough to remedy the adverse environmental impacts of ML. Instead, we argue for systems thinking as the next step towards holistically improving the environmental sustainability of ML.

Tianfang Zhang, Lei Li, Christian Igel et al.

Deep unfolding networks (DUNs) have proven to be a viable approach to compressive sensing (CS). In this work, we propose a DUN called low-rank CS network (LR-CSNet) for natural image CS. Real-world image patches are often well-represented by low-rank approximations. LR-CSNet exploits this property by adding a low-rank prior to the CS optimization task. We derive a corresponding iterative optimization procedure using variable splitting, which is then translated to a new DUN architecture. The architecture uses low-rank generation modules (LRGMs), which learn low-rank matrix factorizations, as well as gradient descent and proximal mappings (GDPMs), which are proposed to extract high-frequency features to refine image details. In addition, the deep features generated at each reconstruction stage in the DUN are transferred between stages to boost the performance. Our extensive experiments on three widely considered datasets demonstrate the promising performance of LR-CSNet compared to state-of-the-art methods in natural image CS.

Lasse Borgholt, Jakob Drachmann Havtorn, Joakim Edin et al.

Unsupervised representation learning for speech processing has matured greatly in the last few years. Work in computer vision and natural language processing has paved the way, but speech data offers unique challenges. As a result, methods from other domains rarely translate directly. We review the development of unsupervised representation learning for speech over the last decade. We identify two primary model categories: self-supervised methods and probabilistic latent variable models. We describe the models and develop a comprehensive taxonomy. Finally, we discuss and compare models from the two categories.

Lei Li, Tianfang Zhang, Stefan Oehmcke et al.

Building segmentation from aerial images and 3D laser scanning (LiDAR) is a challenging task due to the diversity of backgrounds, building textures, and image quality. While current research using different types of convolutional and transformer networks has considerably improved the performance on this task, even more accurate segmentation methods for buildings are desirable for applications such as automatic mapping. In this study, we propose a general framework termed \emph{BuildSeg} employing a generic approach that can be quickly applied to segment buildings. Different data sources were combined to increase generalization performance. The approach yields good results for different data sources as shown by experiments on high-resolution multi-spectral and LiDAR imagery of cities in Norway, Denmark and France. We applied ConvNeXt and SegFormer based models on the high resolution aerial image dataset from the MapAI-competition. The methods achieved an IOU of 0.7902 and a boundary IOU of 0.6185. We used post-processing to account for the rectangular shape of the objects. This increased the boundary IOU from 0.6185 to 0.6189.

Dimitri Gominski, Ankit Kariryaa, Martin Brandt et al.

There is a rising interest in mapping trees using satellite or aerial imagery, but there is no standardized evaluation protocol for comparing and enhancing methods. In dense canopy areas, the high variability of tree sizes and their spatial proximity makes it arduous to define the quality of the predictions. Concurrently, object-centric approaches such as bounding box detection usuallyperform poorly on small and dense objects. It thus remains unclear what is the ideal framework for individual tree mapping, in regards to detection and segmentation approaches, convolutional neural networks and transformers. In this paper, we introduce an evaluation framework suited for individual tree mapping in any physical environment, with annotation costs and applicative goals in mind. We review and compare different approaches and deep architectures, and introduce a new method that we experimentally prove to be a good compromise between segmentation and detection.

Christian Igel

Monotonicity constraints are powerful regularizers in statistical modelling. They can support fairness in computer-aided decision making and increase plausibility in data-driven scientific models. The seminal min-max (MM) neural network architecture ensures monotonicity, but often gets stuck in undesired local optima during training because of partial derivatives of the MM nonlinearities being zero. We propose a simple modification of the MM network using strictly-increasing smooth minimum and maximum functions that alleviates this problem. The resulting smooth min-max (SMM) network module inherits the asymptotic approximation properties from the MM architecture. It can be used within larger deep learning systems trained end-to-end. The SMM module is conceptually simple and computationally less demanding than state-of-the-art neural networks for monotonic modelling. Our experiments show that this does not come with a loss in generalization performance compared to alternative neural and non-neural approaches.

Hui Zhang, Ankit Kariryaa, Venkanna Babu Guthula et al.

Trees inside cities are important for the urban microclimate, contributing positively to the physical and mental health of the urban dwellers. Despite their importance, often only limited information about city trees is available. Therefore in this paper, we propose a method for mapping urban trees in high-resolution aerial imagery using limited datasets and deep learning. Deep learning has become best-practice for this task, however, existing approaches rely on large and accurately labelled training datasets, which can be difficult and expensive to obtain. However, often noisy and incomplete data may be available that can be combined and utilized to solve more difficult tasks than those datasets were intended for. This paper studies how to combine accurate point labels of urban trees along streets with crowd-sourced annotations from an open geographic database to delineate city trees in remote sensing images, a task which is challenging even for humans. To that end, we perform semantic segmentation of very high resolution aerial imagery using a fully convolutional neural network. The main challenge is that our segmentation maps are sparsely annotated and incomplete. Small areas around the point labels of the street trees coming from official and crowd-sourced data are marked as foreground class. Crowd-sourced annotations of streets, buildings, etc. define the background class. Since the tree data is incomplete, we introduce a masking to avoid class confusion. Our experiments in Hamburg, Germany, showed that the system is able to produce tree cover maps, not limited to trees along streets, without providing tree delineations. We evaluated the method on manually labelled trees and show that performance drastically deteriorates if the open geographic database is not used.

Philip Enevoldsen, Christian Gundersen, Nico Lang et al.

Open-set recognition (OSR), the identification of novel categories, can be a critical component when deploying classification models in real-world applications. Recent work has shown that familiarity-based scoring rules such as the Maximum Softmax Probability (MSP) or the Maximum Logit Score (MLS) are strong baselines when the closed-set accuracy is high. However, one of the potential weaknesses of familiarity-based OSR are adversarial attacks. Here, we study gradient-based adversarial attacks on familiarity scores for both types of attacks, False Familiarity and False Novelty attacks, and evaluate their effectiveness in informed and uninformed settings on TinyImageNet. Furthermore, we explore how novel and familiar samples react to adversarial attacks and formulate the adversarial reaction score as an alternative OSR scoring rule, which shows a high correlation with the MLS familiarity score.

Mathias Perslev, Erik Bjørnager Dam, Akshay Pai et al.

Many recent medical segmentation systems rely on powerful deep learning models to solve highly specific tasks. To maximize performance, it is standard practice to evaluate numerous pipelines with varying model topologies, optimization parameters, pre- & postprocessing steps, and even model cascades. It is often not clear how the resulting pipeline transfers to different tasks. We propose a simple and thoroughly evaluated deep learning framework for segmentation of arbitrary medical image volumes. The system requires no task-specific information, no human interaction and is based on a fixed model topology and a fixed hyperparameter set, eliminating the process of model selection and its inherent tendency to cause method-level over-fitting. The system is available in open source and does not require deep learning expertise to use. Without task-specific modifications, the system performed better than or similar to highly specialized deep learning methods across 3 separate segmentation tasks. In addition, it ranked 5-th and 6-th in the first and second round of the 2018 Medical Segmentation Decathlon comprising another 10 tasks. The system relies on multi-planar data augmentation which facilitates the application of a single 2D architecture based on the familiar U-Net. Multi-planar training combines the parameter efficiency of a 2D fully convolutional neural network with a systematic train- and test-time augmentation scheme, which allows the 2D model to learn a representation of the 3D image volume that fosters generalization.

Lucia Gordon, Serge Belongie, Christian Igel et al.

Recent research in geospatial machine learning has demonstrated that models pretrained with self-supervised learning on Earth observation data can perform well on downstream tasks with limited training data. However, most of the existing geospatial benchmark datasets have few data modalities and poor global representation, limiting the ability to evaluate multimodal pretrained models at global scales. To fill this gap, we introduce MMEarth-Bench, a collection of five new multimodal environmental tasks with 12 modalities, globally distributed data, and both in- and out-of-distribution test splits. We benchmark a diverse set of pretrained models and find that while (multimodal) pretraining tends to improve model robustness in limited data settings, geographic generalization abilities remain poor. In order to facilitate model adaptation to new downstream tasks and geographic domains, we propose a model-agnostic method for test-time training with multimodal reconstruction (TTT-MMR) that uses all the modalities available at test time as auxiliary tasks, regardless of whether a pretrained model accepts them as input. Our method improves model performance on both the random and geographic test splits, and geographic batching leads to a good trade-off between regularization and specialization during TTT. Our dataset, code, and visualization tool are linked from the project page at lgordon99.github.io/mmearth-bench.

Sander Riisøen Jyhne, Christian Igel, Morten Goodwin et al.

High-resolution imagery is often hindered by limitations in sensor technology, atmospheric conditions, and costs. Such challenges occur in satellite remote sensing, but also with handheld cameras, such as our smartphones. Hence, super-resolution aims to enhance the image resolution algorithmically. Since single-image super-resolution requires solving an inverse problem, such methods must exploit strong priors, e.g. learned from high-resolution training data, or be constrained by auxiliary data, e.g. by a high-resolution guide from another modality. While qualitatively pleasing, such approaches often lead to "hallucinated" structures that do not match reality. In contrast, multi-image super-resolution (MISR) aims to improve the (optical) resolution by constraining the super-resolution process with multiple views taken with sub-pixel shifts. Here, we propose SuperF, a test-time optimization approach for MISR that leverages coordinate-based neural networks, also called neural fields. Their ability to represent continuous signals with an implicit neural representation (INR) makes them an ideal fit for the MISR task. The key characteristic of our approach is to share an INR for multiple shifted low-resolution frames and to jointly optimize the frame alignment with the INR. Our approach advances related INR baselines, adopted from burst fusion for layer separation, by directly parameterizing the sub-pixel alignment as optimizable affine transformation parameters and by optimizing via a super-sampled coordinate grid that corresponds to the output resolution. Our experiments yield compelling results on simulated bursts of satellite imagery and ground-level images from handheld cameras, with upsampling factors of up to 8. A key advantage of SuperF is that this approach does not rely on any high-resolution training data.

Vishal Nedungadi, Ankit Kariryaa, Stefan Oehmcke et al.

The volume of unlabelled Earth observation (EO) data is huge, but many important applications lack labelled training data. However, EO data offers the unique opportunity to pair data from different modalities and sensors automatically based on geographic location and time, at virtually no human labor cost. We seize this opportunity to create MMEarth, a diverse multi-modal pretraining dataset at global scale. Using this new corpus of 1.2 million locations, we propose a Multi-Pretext Masked Autoencoder (MP-MAE) approach to learn general-purpose representations for optical satellite images. Our approach builds on the ConvNeXt V2 architecture, a fully convolutional masked autoencoder (MAE). Drawing upon a suite of multi-modal pretext tasks, we demonstrate that our MP-MAE approach outperforms both MAEs pretrained on ImageNet and MAEs pretrained on domain-specific satellite images. This is shown on several downstream tasks including image classification and semantic segmentation. We find that pretraining with multi-modal pretext tasks notably improves the linear probing performance compared to pretraining on optical satellite images only. This also leads to better label efficiency and parameter efficiency which are crucial aspects in global scale applications.

Lei Li, Tianfang Zhang, Zhongyu Jiang et al.

Accurate and consistent methods for counting trees based on remote sensing data are needed to support sustainable forest management, assess climate change mitigation strategies, and build trust in tree carbon credits. Two-dimensional remote sensing imagery primarily shows overstory canopy, and it does not facilitate easy differentiation of individual trees in areas with a dense canopy and does not allow for easy separation of trees when the canopy is dense. We leverage the fusion of three-dimensional LiDAR measurements and 2D imagery to facilitate the accurate counting of trees. We compare a deep learning approach to counting trees in forests using 3D airborne LiDAR data and 2D imagery. The approach is compared with state-of-the-art algorithms, like operating on 3D point cloud and 2D imagery. We empirically evaluate the different methods on the NeonTreeCount data set, which we use to define a tree-counting benchmark. The experiments show that FuseCountNet yields more accurate tree counts.

Pedram Bakhtiarifard, Pınar Tözün, Christian Igel et al.

Sustainability encompasses three key facets: economic, environmental, and social. However, the nascent discourse that is emerging on sustainable artificial intelligence (AI) has predominantly focused on the environmental sustainability of AI, often neglecting the economic and social aspects. Achieving truly sustainable AI necessitates addressing the tension between its climate awareness and its social sustainability, which hinges on equitable access to AI development resources. The concept of resource awareness advocates for broader access to the infrastructure required to develop AI, fostering equity in AI innovation. Yet, this push for improving accessibility often overlooks the environmental costs of expanding such resource usage. In this position paper, we argue that reconciling climate and resource awareness is essential to realizing the full potential of sustainable AI. We use the framework of base-superstructure to analyze how the material conditions are influencing the current AI discourse. We also introduce the Climate and Resource Aware Machine Learning (CARAML) framework to address this conflict and propose actionable recommendations spanning individual, community, industry, government, and global levels to achieve sustainable AI.

Lasse Borgholt, Jakob Havtorn, Christian Igel et al.

Modern neural networks have greatly improved performance across speech recognition benchmarks. However, gains are often driven by frequent words with limited semantic weight, which can obscure meaningful differences in word error rate, the primary evaluation metric. Errors in rare terms, named entities, and domain-specific vocabulary are more consequential, but remain hidden by aggregate metrics. This highlights the need for finer-grained error analysis, which depends on accurate alignment between reference and model transcripts. However, conventional alignment methods are not designed for such precision. We propose a novel alignment algorithm that couples dynamic programming with beam search scoring. Compared to traditional text alignment methods, our approach provides more accurate alignment of individual errors, enabling reliable error analysis. The algorithm is made available via PyPI.

Simon B. Jensen, Stefan Oehmcke, Andreas Møgelmose et al.

Assessment of forest biodiversity is crucial for ecosystem management and conservation. While traditional field surveys provide high-quality assessments, they are labor-intensive and spatially limited. This study investigates whether deep learning-based fusion of close-range sensing data from 2D orthophotos and 3D airborne laser scanning (ALS) point clouds can reliable assess the biodiversity potential of forests. We introduce the BioVista dataset, comprising 44 378 paired samples of orthophotos and ALS point clouds from temperate forests in Denmark, designed to explore multimodal fusion approaches. Using deep neural networks (ResNet for orthophotos and PointVector for ALS point clouds), we investigate each data modality's ability to assess forest biodiversity potential, achieving overall accuracies of 76.7% and 75.8%, respectively. We explore various 2D and 3D fusion approaches: confidence-based ensembling, feature-level concatenation, and end-to-end training, achieving overall accuracies of 80.5%, 81.4% and 80.4% respectively. Our results demonstrate that spectral information from orthophotos and structural information from ALS point clouds effectively complement each other in forest biodiversity assessment.

Bob Pepin, Christian Igel, Raghavendra Selvan

We study to which extent additive fairness metrics (statistical parity, equal opportunity and equalized odds) can be influenced in a multi-class classification problem by memorizing a subset of the population. We give explicit expressions for the bias resulting from memorization in terms of the label and group membership distribution of the memorized dataset and the classifier bias on the unmemorized dataset. We also characterize the memorized datasets that eliminate the bias for all three metrics considered. Finally we provide upper and lower bounds on the total probability mass in the memorized dataset that is necessary for the complete elimination of these biases.

Nick Hauptvogel, Christian Igel

It is common practice to combine deep neural networks into ensembles. These deep ensembles can benefit from the cancellation of errors effect: Errors by ensemble members may average out, leading to better generalization performance than each individual network. Bayesian neural networks learn a posterior distribution over model parameters, and sampling and weighting networks according to this posterior yields an ensemble model referred to as a Bayes ensemble. This study reviews the argument that neither the sampling nor the weighting in Bayes ensembles are particularly well suited for increasing generalization performance, as they do not support the cancellation of errors effect. In contrast, we show that a weighted average of models, where the weights are optimized by minimizing a second-order PAC-Bayesian generalization bound, can improve generalization. It is crucial that the optimization takes correlations between models into account. This can be achieved by minimizing the tandem loss, which requires hold-out data for estimating error correlations. The tandem loss based PAC-Bayesian weighting increases robustness against correlated models and models with lower performance in an ensemble. This allows us to safely add several models from the same learning process to an ensemble, instead of using early-stopping for selecting a single weight configuration. Our experiments provide further evidence that state-of-the-art intricate Bayes ensembles do not outperform simple uniformly weighted deep ensembles in terms of classification accuracy. Additionally, we show that these Bayes ensembles cannot match the performance of deep ensembles weighted by optimizing the tandem loss, which additionally provides nonvacuous rigorous generalization guarantees.

Venkanna Babu Guthula, Stefan Oehmcke, Remigio Chilaule et al.

As low-quality housing and in particular certain roof characteristics are associated with an increased risk of malaria, classification of roof types based on remote sensing imagery can support the assessment of malaria risk and thereby help prevent the disease. To support research in this area, we release the Nacala-Roof-Material dataset, which contains high-resolution drone images from Mozambique with corresponding labels delineating houses and specifying their roof types. The dataset defines a multi-task computer vision problem, comprising object detection, classification, and segmentation. In addition, we benchmarked various state-of-the-art approaches on the dataset. Canonical U-Nets, YOLOv8, and a custom decoder on pretrained DINOv2 served as baselines. We show that each of the methods has its advantages but none is superior on all tasks, which highlights the potential of our dataset for future research in multi-task learning. While the tasks are closely related, accurate segmentation of objects does not necessarily imply accurate instance separation, and vice versa. We address this general issue by introducing a variant of the deep ordinal watershed (DOW) approach that additionally separates the interior of objects, allowing for improved object delineation and separation. We show that our DOW variant is a generic approach that improves the performance of both U-Net and DINOv2 backbones, leading to a better trade-off between semantic segmentation and instance segmentation.

Dustin Wright, Christian Igel, Raghavendra Selvan

Modern neural networks are often massively overparameterized leading to high compute costs during training and at inference. One effective method to improve both the compute and energy efficiency of neural networks while maintaining good performance is structured pruning, where full network structures (e.g.~neurons or convolutional filters) that have limited impact on the model output are removed. In this work, we propose Bayesian Model Reduction for Structured pruning (BMRS), a fully end-to-end Bayesian method of structured pruning. BMRS is based on two recent methods: Bayesian structured pruning with multiplicative noise, and Bayesian model reduction (BMR), a method which allows efficient comparison of Bayesian models under a change in prior. We present two realizations of BMRS derived from different priors which yield different structured pruning characteristics: 1) BMRS_N with the truncated log-normal prior, which offers reliable compression rates and accuracy without the need for tuning any thresholds and 2) BMRS_U with the truncated log-uniform prior that can achieve more aggressive compression based on the boundaries of truncation. Overall, we find that BMRS offers a theoretically grounded approach to structured pruning of neural networks yielding both high compression rates and accuracy. Experiments on multiple datasets and neural networks of varying complexity showed that the two BMRS methods offer a competitive performance-efficiency trade-off compared to other pruning methods.

Raghavendra Selvan, Bob Pepin, Christian Igel et al.

The recent advances in deep learning (DL) have been accelerated by access to large-scale data and compute. These large-scale resources have been used to train progressively larger models which are resource intensive in terms of compute, data, energy, and carbon emissions. These costs are becoming a new type of entry barrier to researchers and practitioners with limited access to resources at such scale, particularly in the Global South. In this work, we take a comprehensive look at the landscape of existing DL models for medical image analysis tasks and demonstrate their usefulness in settings where resources are limited. To account for the resource consumption of DL models, we introduce a novel measure to estimate the performance per resource unit, which we call the PePR score. Using a diverse family of 131 unique DL architectures (spanning 1M to 130M trainable parameters) and three medical image datasets, we capture trends about the performance-resource trade-offs. In applications like medical image analysis, we argue that small-scale, specialized models are better than striving for large-scale models. Furthermore, we show that using existing pretrained models that are fine-tuned on new data can significantly reduce the computational resources and data required compared to training models from scratch. We hope this work will encourage the community to focus on improving AI equity by developing methods and models with smaller resource footprints.

Christian Igel, Stefan Oehmcke

When training deep learning models for least-squares regression, we cannot expect that the training error residuals of the final model, selected after a fixed training time or based on performance on a hold-out data set, sum to zero. This can introduce a systematic error that accumulates if we are interested in the total aggregated performance over many data points. We suggest to adjust the bias of the machine learning model after training as a default postprocessing step, which efficiently solves the problem. The severeness of the error accumulation and the effectiveness of the bias correction is demonstrated in exemplary experiments.

Stefan Oehmcke, Lei Li, Katerina Trepekli et al.

Quantification of forest biomass stocks and their dynamics is important for implementing effective climate change mitigation measures. The knowledge is needed, e.g., for local forest management, studying the processes driving af-, re-, and deforestation, and can improve the accuracy of carbon-accounting. Remote sensing using airborne LiDAR can be used to perform these measurements of vegetation structure at large scale. We present deep learning systems for predicting wood volume, above-ground biomass (AGB), and subsequently above-ground carbon stocks directly from airborne LiDAR point clouds. We devise different neural network architectures for point cloud regression and evaluate them on remote sensing data of areas for which AGB estimates have been obtained from field measurements in the Danish national forest inventory. Our adaptation of Minkowski convolutional neural networks for regression gave the best results. The deep neural networks produced significantly more accurate wood volume, AGB, and carbon stock estimates compared to state-of-the-art approaches operating on basic statistics of the point clouds. In contrast to other methods, the proposed deep learning approach does not require a digital terrain model. We expect this finding to have a strong impact on LiDAR-based analyses of biomass dynamics.

Lasse Borgholt, Jakob Drachmann Havtorn, Mostafa Abdou et al.

Spoken language understanding (SLU) tasks are usually solved by first transcribing an utterance with automatic speech recognition (ASR) and then feeding the output to a text-based model. Recent advances in self-supervised representation learning for speech data have focused on improving the ASR component. We investigate whether representation learning for speech has matured enough to replace ASR in SLU. We compare learned speech features from wav2vec 2.0, state-of-the-art ASR transcripts, and the ground truth text as input for a novel speech-based named entity recognition task, a cardiac arrest detection task on real-world emergency calls and two existing SLU benchmarks. We show that learned speech features are superior to ASR transcripts on three classification tasks. For machine translation, ASR transcripts are still the better choice. We highlight the intrinsic robustness of wav2vec 2.0 representations to out-of-vocabulary words as key to better performance.

Yi-Shan Wu, Andrés R. Masegosa, Stephan S. Lorenzen et al.

We present a new second-order oracle bound for the expected risk of a weighted majority vote. The bound is based on a novel parametric form of the Chebyshev- Cantelli inequality (a.k.a. one-sided Chebyshev's), which is amenable to efficient minimization. The new form resolves the optimization challenge faced by prior oracle bounds based on the Chebyshev-Cantelli inequality, the C-bounds [Germain et al., 2015], and, at the same time, it improves on the oracle bound based on second order Markov's inequality introduced by Masegosa et al. [2020]. We also derive a new concentration of measure inequality, which we name PAC-Bayes-Bennett, since it combines PAC-Bayesian bounding with Bennett's inequality. We use it for empirical estimation of the oracle bound. The PAC-Bayes-Bennett inequality improves on the PAC-Bayes-Bernstein inequality of Seldin et al. [2012]. We provide an empirical evaluation demonstrating that the new bounds can improve on the work of Masegosa et al. [2020]. Both the parametric form of the Chebyshev-Cantelli inequality and the PAC-Bayes-Bennett inequality may be of independent interest for the study of concentration of measure in other domains.

Stephan Sloth Lorenzen, Christian Igel, Mads Nielsen

The information bottleneck (IB) principle has been suggested as a way to analyze deep neural networks. The learning dynamics are studied by inspecting the mutual information (MI) between the hidden layers and the input and output. Notably, separate fitting and compression phases during training have been reported. This led to some controversy including claims that the observations are not reproducible and strongly dependent on the type of activation function used as well as on the way the MI is estimated. Our study confirms that different ways of binning when computing the MI lead to qualitatively different results, either supporting or refusing IB conjectures. To resolve the controversy, we study the IB principle in settings where MI is non-trivial and can be computed exactly. We monitor the dynamics of quantized neural networks, that is, we discretize the whole deep learning system so that no approximation is required when computing the MI. This allows us to quantify the information flow without measurement errors. In this setting, we observed a fitting phase for all layers and a compression phase for the output layer in all experiments; the compression in the hidden layers was dependent on the type of activation function. Our study shows that the initial IB results were not artifacts of binning when computing the MI. However, the critical claim that the compression phase may not be observed for some networks also holds true.

Lasse Borgholt, Jakob Drachmann Havtorn, Željko Agić et al.

The two most common paradigms for end-to-end speech recognition are connectionist temporal classification (CTC) and attention-based encoder-decoder (AED) models. It has been argued that the latter is better suited for learning an implicit language model. We test this hypothesis by measuring temporal context sensitivity and evaluate how the models perform when we constrain the amount of contextual information in the audio input. We find that the AED model is indeed more context sensitive, but that the gap can be closed by adding self-attention to the CTC model. Furthermore, the two models perform similarly when contextual information is constrained. Finally, in contrast to previous research, our results show that the CTC model is highly competitive on WSJ and LibriSpeech without the help of an external language model.

Lasse Borgholt, Tycho Max Sylvester Tax, Jakob Drachmann Havtorn et al.

Recent advances in self-supervised learning through contrastive training have shown that it is possible to learn a competitive speech recognition system with as little as 10 minutes of labeled data. However, these systems are computationally expensive since they require pre-training followed by fine-tuning in a large parameter space. We explore the performance of such systems without fine-tuning by training a state-of-the-art speech recognizer on the fixed representations from the computationally demanding wav2vec 2.0 framework. We find performance to decrease without fine-tuning and, in the extreme low-resource setting, wav2vec 2.0 is inferior to its predecessor. In addition, we find that wav2vec 2.0 representations live in a low dimensional subspace and that decorrelating the features of the representations can stabilize training of the automatic speech recognizer. Finally, we propose a bidirectional extension to the original wav2vec framework that consistently improves performance.

Svetlana Kutuzova, Oswin Krause, Douglas McCloskey et al.

Multimodal generative models should be able to learn a meaningful latent representation that enables a coherent joint generation of all modalities (e.g., images and text). Many applications also require the ability to accurately sample modalities conditioned on observations of a subset of the modalities. Often not all modalities may be observed for all training data points, so semi-supervised learning should be possible. In this study, we propose a novel product-of-experts (PoE) based variational autoencoder that have these desired properties. We benchmark it against a mixture-of-experts (MoE) approach and an approach of combining the modalities with an additional encoder network. An empirical evaluation shows that the PoE based models can outperform the contrasted models. Our experiments support the intuition that PoE models are more suited for a conjunctive combination of modalities.

Andrés R. Masegosa, Stephan S. Lorenzen, Christian Igel et al.

We present a novel analysis of the expected risk of weighted majority vote in multiclass classification. The analysis takes correlation of predictions by ensemble members into account and provides a bound that is amenable to efficient minimization, which yields improved weighting for the majority vote. We also provide a specialized version of our bound for binary classification, which allows to exploit additional unlabeled data for tighter risk estimation. In experiments, we apply the bound to improve weighting of trees in random forests and show that, in contrast to the commonly used first order bound, minimization of the new bound typically does not lead to degradation of the test error of the ensemble.

Steffen Czolbe, Oswin Krause, Ingemar Cox et al.

To train Variational Autoencoders (VAEs) to generate realistic imagery requires a loss function that reflects human perception of image similarity. We propose such a loss function based on Watson's perceptual model, which computes a weighted distance in frequency space and accounts for luminance and contrast masking. We extend the model to color images, increase its robustness to translation by using the Fourier Transform, remove artifacts due to splitting the image into blocks, and make it differentiable. In experiments, VAEs trained with the new loss function generated realistic, high-quality image samples. Compared to using the Euclidean distance and the Structural Similarity Index, the images were less blurry; compared to deep neural network based losses, the new approach required less computational resources and generated images with less artifacts.

Kai Brügge, Asja Fischer, Christian Igel

The Metropolis algorithm is arguably the most fundamental Markov chain Monte Carlo (MCMC) method. But the algorithm is not guaranteed to converge to the desired distribution in the case of multivariate binary distributions (e.g., Ising models or stochastic neural networks such as Boltzmann machines) if the variables (sites or neurons) are updated in a fixed order, a setting commonly used in practice. The reason is that the corresponding Markov chain may not be irreducible. We propose a modified Metropolis transition operator that behaves almost always identically to the standard Metropolis operator and prove that it ensures irreducibility and convergence to the limiting distribution in the multivariate binary case with fixed-order updates. The result provides an explanation for the behaviour of Metropolis MCMC in that setting and closes a long-standing theoretical gap. We experimentally studied the standard and modified Metropolis operator for models were they actually behave differently. If the standard algorithm also converges, the modified operator exhibits similar (if not better) performance in terms of convergence speed.

Arjun D. Desai, Francesco Caliva, Claudia Iriondo et al.

Purpose: To organize a knee MRI segmentation challenge for characterizing the semantic and clinical efficacy of automatic segmentation methods relevant for monitoring osteoarthritis progression. Methods: A dataset partition consisting of 3D knee MRI from 88 subjects at two timepoints with ground-truth articular (femoral, tibial, patellar) cartilage and meniscus segmentations was standardized. Challenge submissions and a majority-vote ensemble were evaluated using Dice score, average symmetric surface distance, volumetric overlap error, and coefficient of variation on a hold-out test set. Similarities in network segmentations were evaluated using pairwise Dice correlations. Articular cartilage thickness was computed per-scan and longitudinally. Correlation between thickness error and segmentation metrics was measured using Pearson's coefficient. Two empirical upper bounds for ensemble performance were computed using combinations of model outputs that consolidated true positives and true negatives. Results: Six teams (T1-T6) submitted entries for the challenge. No significant differences were observed across all segmentation metrics for all tissues (p=1.0) among the four top-performing networks (T2, T3, T4, T6). Dice correlations between network pairs were high (>0.85). Per-scan thickness errors were negligible among T1-T4 (p=0.99) and longitudinal changes showed minimal bias (<0.03mm). Low correlations (<0.41) were observed between segmentation metrics and thickness error. The majority-vote ensemble was comparable to top performing networks (p=1.0). Empirical upper bound performances were similar for both combinations (p=1.0). Conclusion: Diverse networks learned to segment the knee similarly where high segmentation accuracy did not correlate to cartilage thickness accuracy. Voting ensembles did not outperform individual networks but may help regularize individual models.

Urun Dogan, Aniket Anand Deshmukh, Marcin Machura et al.

Curriculum learning can improve neural network training by guiding the optimization to desirable optima. We propose a novel curriculum learning approach for image classification that adapts the loss function by changing the label representation. The idea is to use a probability distribution over classes as target label, where the class probabilities reflect the similarity to the true class. Gradually, this label representation is shifted towards the standard one-hot-encoding. That is, in the beginning minor mistakes are corrected less than large mistakes, resembling a teaching process in which broad concepts are explained first before subtle differences are taught. The class similarity can be based on prior knowledge. For the special case of the labels being natural words, we propose a generic way to automatically compute the similarities. The natural words are embedded into Euclidean space using a standard word embedding. The probability of each class is then a function of the cosine similarity between the vector representations of the class and the true label. The proposed label-similarity curriculum learning (LCL) approach was empirically evaluated using several popular deep learning architectures for image classification tasks applied to five datasets including ImageNet, CIFAR100, and AWA2. In all scenarios, LCL was able to improve the classification accuracy on the test data compared to standard training.

Mathias Perslev, Michael Hejselbak Jensen, Sune Darkner et al.

Neural networks are becoming more and more popular for the analysis of physiological time-series. The most successful deep learning systems in this domain combine convolutional and recurrent layers to extract useful features to model temporal relations. Unfortunately, these recurrent models are difficult to tune and optimize. In our experience, they often require task-specific modifications, which makes them challenging to use for non-experts. We propose U-Time, a fully feed-forward deep learning approach to physiological time series segmentation developed for the analysis of sleep data. U-Time is a temporal fully convolutional network based on the U-Net architecture that was originally proposed for image segmentation. U-Time maps sequential inputs of arbitrary length to sequences of class labels on a freely chosen temporal scale. This is done by implicitly classifying every individual time-point of the input signal and aggregating these classifications over fixed intervals to form the final predictions. We evaluated U-Time for sleep stage classification on a large collection of sleep electroencephalography (EEG) datasets. In all cases, we found that U-Time reaches or outperforms current state-of-the-art deep learning models while being much more robust in the training process and without requiring architecture or hyperparameter adaptation across tasks.

Mauricio Orbes-Arteaga, Jorge Cardoso, Lauge Sørensen et al.

In the absence of sufficient data variation (e.g., scanner and protocol variability) in annotated data, deep neural networks (DNNs) tend to overfit during training. As a result, their performance is significantly lower on data from unseen sources compared to the performance on data from the same source as the training data. Semi-supervised domain adaptation methods can alleviate this problem by tuning networks to new target domains without the need for annotated data from these domains. Adversarial domain adaptation (ADA) methods are a popular choice that aim to train networks in such a way that the features generated are domain agnostic. However, these methods require careful dataset-specific selection of hyperparameters such as the complexity of the discriminator in order to achieve a reasonable performance. We propose to use knowledge distillation (KD) -- an efficient way of transferring knowledge between different DNNs -- for semi-supervised domain adaption of DNNs. It does not require dataset-specific hyperparameter tuning, making it generally applicable. The proposed method is compared to ADA for segmentation of white matter hyperintensities (WMH) in magnetic resonance imaging (MRI) scans generated by scanners that are not a part of the training set. Compared with both the baseline DNN (trained on source domain only and without any adaption to target domain) and with using ADA for semi-supervised domain adaptation, the proposed method achieves significantly higher WMH dice scores.

Patrick Bilic, Patrick Christ, Hongwei Bran Li et al.

In this work, we report the set-up and results of the Liver Tumor Segmentation Benchmark (LiTS), which was organized in conjunction with the IEEE International Symposium on Biomedical Imaging (ISBI) 2017 and the International Conferences on Medical Image Computing and Computer-Assisted Intervention (MICCAI) 2017 and 2018. The image dataset is diverse and contains primary and secondary tumors with varied sizes and appearances with various lesion-to-background levels (hyper-/hypo-dense), created in collaboration with seven hospitals and research institutions. Seventy-five submitted liver and liver tumor segmentation algorithms were trained on a set of 131 computed tomography (CT) volumes and were tested on 70 unseen test images acquired from different patients. We found that not a single algorithm performed best for both liver and liver tumors in the three events. The best liver segmentation algorithm achieved a Dice score of 0.963, whereas, for tumor segmentation, the best algorithms achieved Dices scores of 0.674 (ISBI 2017), 0.702 (MICCAI 2017), and 0.739 (MICCAI 2018). Retrospectively, we performed additional analysis on liver tumor detection and revealed that not all top-performing segmentation algorithms worked well for tumor detection. The best liver tumor detection method achieved a lesion-wise recall of 0.458 (ISBI 2017), 0.515 (MICCAI 2017), and 0.554 (MICCAI 2018), indicating the need for further research. LiTS remains an active benchmark and resource for research, e.g., contributing the liver-related segmentation tasks in \url{http://medicaldecathlon.com/}. In addition, both data and online evaluation are accessible via \url{www.lits-challenge.com}.

Stephan Sloth Lorenzen, Christian Igel, Yevgeny Seldin

Existing guarantees in terms of rigorous upper bounds on the generalization error for the original random forest algorithm, one of the most frequently used machine learning methods, are unsatisfying. We discuss and evaluate various PAC-Bayesian approaches to derive such bounds. The bounds do not require additional hold-out data, because the out-of-bag samples from the bagging in the training process can be exploited. A random forest predicts by taking a majority vote of an ensemble of decision trees. The first approach is to bound the error of the vote by twice the error of the corresponding Gibbs classifier (classifying with a single member of the ensemble selected at random). However, this approach does not take into account the effect of averaging out of errors of individual classifiers when taking the majority vote. This effect provides a significant boost in performance when the errors are independent or negatively correlated, but when the correlations are strong the advantage from taking the majority vote is small. The second approach based on PAC-Bayesian C-bounds takes dependencies between ensemble members into account, but it requires estimating correlations between the errors of the individual classifiers. When the correlations are high or the estimation is poor, the bounds degrade. In our experiments, we compute generalization bounds for random forests on various benchmark data sets. Because the individual decision trees already perform well, their predictions are highly correlated and the C-bounds do not lead to satisfactory results. For the same reason, the bounds based on the analysis of Gibbs classifiers are typically superior and often reasonably tight. Bounds based on a validation set coming at the cost of a smaller training set gave better performance guarantees, but worse performance in most experiments.

Mauricio Orbes Arteaga, Lauge Sørensen, M. Jorge Cardoso et al.

For proper generalization performance of convolutional neural networks (CNNs) in medical image segmentation, the learnt features should be invariant under particular non-linear shape variations of the input. To induce invariance in CNNs to such transformations, we propose Probabilistic Augmentation of Data using Diffeomorphic Image Transformation (PADDIT) -- a systematic framework for generating realistic transformations that can be used to augment data for training CNNs. We show that CNNs trained with PADDIT outperforms CNNs trained without augmentation and with generic augmentation in segmenting white matter hyperintensities from T1 and FLAIR brain MRI scans.

Fabian Gieseke, Christian Igel

Without access to large compute clusters, building random forests on large datasets is still a challenging problem. This is, in particular, the case if fully-grown trees are desired. We propose a simple yet effective framework that allows to efficiently construct ensembles of huge trees for hundreds of millions or even billions of training instances using a cheap desktop computer with commodity hardware. The basic idea is to consider a multi-level construction scheme, which builds top trees for small random subsets of the available data and which subsequently distributes all training instances to the top trees' leaves for further processing. While being conceptually simple, the overall efficiency crucially depends on the particular implementation of the different phases. The practical merits of our approach are demonstrated using dense datasets with hundreds of millions of training instances.

Jan Kremer, Kristoffer Stensbo-Smidt, Fabian Gieseke et al.

Astrophysics and cosmology are rich with data. The advent of wide-area digital cameras on large aperture telescopes has led to ever more ambitious surveys of the sky. Data volumes of entire surveys a decade ago can now be acquired in a single night and real-time analysis is often desired. Thus, modern astronomy requires big data know-how, in particular it demands highly efficient machine learning and image analysis algorithms. But scalability is not the only challenge: Astronomy applications touch several current machine learning research questions, such as learning from biased data and dealing with label and measurement noise. We argue that this makes astronomy a great domain for computer science research, as it pushes the boundaries of data analysis. In the following, we will present this exciting application area for data scientists. We will focus on exemplary results, discuss main challenges, and highlight some recent methodological advancements in machine learning and image analysis triggered by astronomical applications.

Niklas Thiemann, Christian Igel, Olivier Wintenberger et al.

We propose a new PAC-Bayesian bound and a way of constructing a hypothesis space, so that the bound is convex in the posterior distribution and also convex in a trade-off parameter between empirical performance of the posterior distribution and its complexity. The complexity is measured by the Kullback-Leibler divergence to a prior. We derive an alternating procedure for minimizing the bound. We show that the bound can be rewritten as a one-dimensional function of the trade-off parameter and provide sufficient conditions under which the function has a single global minimum. When the conditions are satisfied the alternating minimization is guaranteed to converge to the global minimum of the bound. We provide experimental results demonstrating that rigorous minimization of the bound is competitive with cross-validation in tuning the trade-off between complexity and empirical performance. In all our experiments the trade-off turned to be quasiconvex even when the sufficient conditions were violated.

Fabian Gieseke, Cosmin Eugen Oancea, Ashish Mahabal et al.

A buffer k-d tree is a k-d tree variant for massively-parallel nearest neighbor search. While providing valuable speed-ups on modern many-core devices in case both a large number of reference and query points are given, buffer k-d trees are limited by the amount of points that can fit on a single device. In this work, we show how to modify the original data structure and the associated workflow to make the overall approach capable of dealing with massive data sets. We further provide a simple yet efficient way of using multiple devices given in a single workstation. The applicability of the modified framework is demonstrated in the context of astronomy, a field that is faced with huge amounts of data.

Kristoffer Stensbo-Smidt, Fabian Gieseke, Christian Igel et al.

Large-scale surveys make huge amounts of photometric data available. Because of the sheer amount of objects, spectral data cannot be obtained for all of them. Therefore it is important to devise techniques for reliably estimating physical properties of objects from photometric information alone. These estimates are needed to automatically identify interesting objects worth a follow-up investigation as well as to produce the required data for a statistical analysis of the space covered by a survey. We argue that machine learning techniques are suitable to compute these estimates accurately and efficiently. This study promotes a feature selection algorithm, which selects the most informative magnitudes and colours for a given task of estimating physical quantities from photometric data alone. Using k nearest neighbours regression, a well-known non-parametric machine learning method, we show that using the found features significantly increases the accuracy of the estimations compared to using standard features and standard methods. We illustrate the usefulness of the approach by estimating specific star formation rates (sSFRs) and redshifts (photo-z's) using only the broad-band photometry from the Sloan Digital Sky Survey (SDSS). For estimating sSFRs, we demonstrate that our method produces better estimates than traditional spectral energy distribution (SED) fitting. For estimating photo-z's, we show that our method produces more accurate photo-z's than the method employed by SDSS. The study highlights the general importance of performing proper model selection to improve the results of machine learning systems and how feature selection can provide insights into the predictive relevance of particular input features.

Oswin Krause, Asja Fischer, Christian Igel

Estimating the log-likelihood gradient with respect to the parameters of a Restricted Boltzmann Machine (RBM) typically requires sampling using Markov Chain Monte Carlo (MCMC) techniques. To save computation time, the Markov chains are only run for a small number of steps, which leads to a biased estimate. This bias can cause RBM training algorithms such as Contrastive Divergence (CD) learning to deteriorate. We adopt the idea behind Population Monte Carlo (PMC) methods to devise a new RBM training algorithm termed Population-Contrastive-Divergence (pop-CD). Compared to CD, it leads to a consistent estimate and may have a significantly lower bias. Its computational overhead is negligible compared to CD. However, the variance of the gradient estimate increases. We experimentally show that pop-CD can significantly outperform CD. In many cases, we observed a smaller bias and achieved higher log-likelihood values. However, when the RBM distribution has many hidden neurons, the consistent estimate of pop-CD may still have a considerable bias and the variance of the gradient estimate requires a smaller learning rate. Thus, despite its superior theoretical properties, it is not advisable to use pop-CD in its current form on large problems.