Mehmet Ozgur Turkoglu

Mehmet Ozgur Turkoglu, Alexander Becker, Hüseyin Anil Gündüz et al.

The ability to estimate epistemic uncertainty is often crucial when deploying machine learning in the real world, but modern methods often produce overconfident, uncalibrated uncertainty predictions. A common approach to quantify epistemic uncertainty, usable across a wide class of prediction models, is to train a model ensemble. In a naive implementation, the ensemble approach has high computational cost and high memory demand. This challenges in particular modern deep learning, where even a single deep network is already demanding in terms of compute and memory, and has given rise to a number of attempts to emulate the model ensemble without actually instantiating separate ensemble members. We introduce FiLM-Ensemble, a deep, implicit ensemble method based on the concept of Feature-wise Linear Modulation (FiLM). That technique was originally developed for multi-task learning, with the aim of decoupling different tasks. We show that the idea can be extended to uncertainty quantification: by modulating the network activations of a single deep network with FiLM, one obtains a model ensemble with high diversity, and consequently well-calibrated estimates of epistemic uncertainty, with low computational overhead in comparison. Empirically, FiLM-Ensemble outperforms other implicit ensemble methods, and it and comes very close to the upper bound of an explicit ensemble of networks (sometimes even beating it), at a fraction of the memory cost.

Fabian Dionys Schrag, Mehmet Ozgur Turkoglu, Konrad Schindler et al.

Domain adaptation (DA) addresses the challenge of transferring a machine learning model trained on a source domain to a target domain with a different data distribution. In this work, we study DA for the task of Rumex obtusifolius (Rumex) image classification. We train models on a published, ground vehicle-based dataset (source) and evaluate their performance on a custom target dataset acquired by unmanned aerial vehicles (UAVs). We find that Convolutional Neural Network (CNN) models, specifically ResNets, generalize poorly to the target domain, even after fine-tuning on the source data. Applying moment-matching and maximum classifier discrepancy, two established DA techniques, substantially improves target-domain performance. However, Vision Transformer (ViT) models pretrained with self-supervised objectives (DINOv2, DINOv3) handle domain shifts intrinsically well, surpassing even moment-matching-trained ResNets, likely due to the rich, general-purpose representations acquired during large-scale pretraining. Using ViTs fine-tuned on the source dataset, we demonstrate high classification performances in the range of F1=0.8 on our target dataset. To support further research on DA for weed detection in grassland systems, we publicly release our UAV-based target dataset AGSMultiRumex, comprising data from 15 flights over Swiss meadows.

Mehmet Ozgur Turkoglu, Dominik J. Mühlematter, Alexander Becker et al.

Foundation models have become a dominant paradigm in machine learning, achieving remarkable performance across diverse tasks through large-scale pretraining. However, these models often yield overconfident, uncalibrated predictions. The standard approach to quantifying epistemic uncertainty, training an ensemble of independent models, incurs prohibitive computational costs that scale linearly with ensemble size, making it impractical for large foundation models. We propose Singular Value Ensemble (SVE), a parameter-efficient implicit ensemble method that builds on a simple, but powerful core assumption: namely, that the singular vectors of the weight matrices constitute meaningful subspaces of the model's knowledge. Pretrained foundation models encode rich, transferable information in their weight matrices. If the singular vectors are indeed meaningful (orthogonal) "knowledge directions". To obtain a model ensemble, we modulate only how strongly each direction contributes to the output. Rather than learning entirely new parameters, we freeze the singular vectors and only train per-member singular values that rescale the contribution of each direction in that shared knowledge basis. Ensemble diversity emerges naturally as stochastic initialization and random sampling of mini-batches during joint training cause different members to converge to different combinations of the same underlying knowledge. SVE achieves uncertainty quantification comparable to explicit deep ensembles while increasing the parameter count of the base model by less than 1%, making principled uncertainty estimation accessible in resource-constrained settings. We validate SVE on NLP and vision tasks with various different backbones and show that it improves calibration while maintaining predictive accuracy.

Dominik J. Mühlematter, Michelle Halbheer, Alexander Becker et al.

Numerous real-world decisions rely on machine learning algorithms and require calibrated uncertainty estimates. However, modern methods often yield overconfident, uncalibrated predictions. The dominant approach to quantifying the uncertainty inherent in the model is to train an ensemble of separate predictors and measure their empirical variance. In an explicit implementation, the ensemble has high computational cost and memory footprint, especially if the base model itself is already large, like modern transformers. This motivates efforts to develop implicit ensemble methods that emulate the ensemble without explicitly instantiating all its members. We introduce LoRA-Ensemble, a parameter-efficient ensembling method for self-attention networks. It is based on Low-Rank Adaptation (LoRA), originally developed for efficient LLM fine-tuning, and extends it into an implicit ensembling scheme, where all ensemble members share the same, pre-trained self-attention network, but have individual low-rank matrices for the attention projections. The resulting method not only outperforms state-of-the-art implicit techniques like BatchEnsemble, but even matches or exceeds the accuracy of an Explicit Ensemble, while at the same time achieving superior calibration.

Mehmet Ozgur Turkoglu, Selene Ledain, Helge Aasen

Crop type classification using optical satellite time series remains limited in its ability to generalize across seasons, particularly when crop phenology shifts due to inter-annual weather variability. This hampers real-world applicability in scenarios where current-year labels are unavailable. In addition, uncertainty quantification is often overlooked, which reduces the reliability of such approaches for operational crop monitoring. Inspired by ecophysiological principles of plant growth, we propose a simple, model-agnostic Thermal-Time-based Temporal Sampling (T3S) method that replaces calendar time with thermal time. By subsampling time series in this biologically meaningful way, our method highlights key periods within the growing season while reducing temporal redundancy and noise. We evaluate the T3S on a multi-year Sentinel-2 dataset covering the entirety of Switzerland, which allows us to assess all applied methods on unseen years. Compared to state-of-the-art baselines, our approach yields substantial improvements in classification accuracy and, critically, provides well-calibrated uncertainty estimates. Moreover, the T3S method excels in low-data regimes and enables significantly more accurate early-season classification. With just 10% of the training labels, it outperforms the current baseline in both accuracy and uncertainty calibration, and by the end of June, it achieves a performance similar to the full-season baseline model.

Mehmet Ozgur Turkoglu, Eric Brachmann, Konrad Schindler et al.

Visual re-localization means using a single image as input to estimate the camera's location and orientation relative to a pre-recorded environment. The highest-scoring methods are "structure based," and need the query camera's intrinsics as an input to the model, with careful geometric optimization. When intrinsics are absent, methods vie for accuracy by making various other assumptions. This yields fairly good localization scores, but the models are "narrow" in some way, eg., requiring costly test-time computations, or depth sensors, or multiple query frames. In contrast, our proposed method makes few special assumptions, and is fairly lightweight in training and testing. Our pose regression network learns from only relative poses of training scenes. For inference, it builds a graph connecting the query image to training counterparts and uses a graph neural network (GNN) with image representations on nodes and image-pair representations on edges. By efficiently passing messages between them, both representation types are refined to produce a consistent camera pose estimate. We validate the effectiveness of our approach on both standard indoor (7-Scenes) and outdoor (Cambridge Landmarks) camera re-localization benchmarks. Our relative pose regression method matches the accuracy of absolute pose regression networks, while retaining the relative-pose models' test-time speed and ability to generalize to non-training scenes.

Mehmet Ozgur Turkoglu, Stefano D'Aronco, Gregor Perich et al.

The aim of this paper is to map agricultural crops by classifying satellite image time series. Domain experts in agriculture work with crop type labels that are organised in a hierarchical tree structure, where coarse classes (like orchards) are subdivided into finer ones (like apples, pears, vines, etc.). We develop a crop classification method that exploits this expert knowledge and significantly improves the mapping of rare crop types. The three-level label hierarchy is encoded in a convolutional, recurrent neural network (convRNN), such that for each pixel the model predicts three labels at different level of granularity. This end-to-end trainable, hierarchical network architecture allows the model to learn joint feature representations of rare classes (e.g., apples, pears) at a coarser level (e.g., orchard), thereby boosting classification performance at the fine-grained level. Additionally, labelling at different granularity also makes it possible to adjust the output according to the classification scores; as coarser labels with high confidence are sometimes more useful for agricultural practice than fine-grained but very uncertain labels. We validate the proposed method on a new, large dataset that we make public. ZueriCrop covers an area of 50 km x 48 km in the Swiss cantons of Zurich and Thurgau with a total of 116'000 individual fields spanning 48 crop classes, and 28,000 (multi-temporal) image patches from Sentinel-2. We compare our proposed hierarchical convRNN model with several baselines, including methods designed for imbalanced class distributions. The hierarchical approach performs superior by at least 9.9 percentage points in F1-score.

Nando Metzger, Mehmet Ozgur Turkoglu, Stefano D'Aronco et al.

Optical satellite sensors cannot see the Earth's surface through clouds. Despite the periodic revisit cycle, image sequences acquired by Earth observation satellites are therefore irregularly sampled in time. State-of-the-art methods for crop classification (and other time series analysis tasks) rely on techniques that implicitly assume regular temporal spacing between observations, such as recurrent neural networks (RNNs). We propose to use neural ordinary differential equations (NODEs) in combination with RNNs to classify crop types in irregularly spaced image sequences. The resulting ODE-RNN models consist of two steps: an update step, where a recurrent unit assimilates new input data into the model's hidden state; and a prediction step, in which NODE propagates the hidden state until the next observation arrives. The prediction step is based on a continuous representation of the latent dynamics, which has several advantages. At the conceptual level, it is a more natural way to describe the mechanisms that govern the phenological cycle. From a practical point of view, it makes it possible to sample the system state at arbitrary points in time, such that one can integrate observations whenever they are available, and extrapolate beyond the last observation. Our experiments show that ODE-RNN indeed improves classification accuracy over common baselines such as LSTM, GRU, and temporal convolution. The gains are most prominent in the challenging scenario where only few observations are available (i.e., frequent cloud cover). Moreover, we show that the ability to extrapolate translates to better classification performance early in the season, which is important for forecasting.

Mehmet Ozgur Turkoglu, Stefano D'Aronco, Jan Dirk Wegner et al.

We propose a new STAckable Recurrent cell (STAR) for recurrent neural networks (RNNs), which has fewer parameters than widely used LSTM and GRU while being more robust against vanishing or exploding gradients. Stacking recurrent units into deep architectures suffers from two major limitations: (i) many recurrent cells (e.g., LSTMs) are costly in terms of parameters and computation resources; and (ii) deep RNNs are prone to vanishing or exploding gradients during training. We investigate the training of multi-layer RNNs and examine the magnitude of the gradients as they propagate through the network in the "vertical" direction. We show that, depending on the structure of the basic recurrent unit, the gradients are systematically attenuated or amplified. Based on our analysis we design a new type of gated cell that better preserves gradient magnitude. We validate our design on a large number of sequence modelling tasks and demonstrate that the proposed STAR cell allows to build and train deeper recurrent architectures, ultimately leading to improved performance while being computationally more efficient.

Mehmet Ozgur Turkoglu, William Thong, Luuk Spreeuwers et al.

The visual world we sense, interpret and interact everyday is a complex composition of interleaved physical entities. Therefore, it is a very challenging task to generate vivid scenes of similar complexity using computers. In this work, we present a scene generation framework based on Generative Adversarial Networks (GANs) to sequentially compose a scene, breaking down the underlying problem into smaller ones. Different than the existing approaches, our framework offers an explicit control over the elements of a scene through separate background and foreground generators. Starting with an initially generated background, foreground objects then populate the scene one-by-one in a sequential manner. Via quantitative and qualitative experiments on a subset of the MS-COCO dataset, we show that our proposed framework produces not only more diverse images but also copes better with affine transformations and occlusion artifacts of foreground objects than its counterparts.