Sam Leroux

Sam Leroux, Pieter Simoens, Meelis Lootus et al.

Deploying machine learning applications on edge devices can bring clear benefits such as improved reliability, latency and privacy but it also introduces its own set of challenges. Most works focus on the limited computational resources of edge platforms but this is not the only bottleneck standing in the way of widespread adoption. In this paper we list several other challenges that a TinyML practitioner might need to consider when operationalizing an application on edge devices. We focus on tasks such as monitoring and managing the application, common functionality for a MLOps platform, and show how they are complicated by the distributed nature of edge deployment. We also discuss issues that are unique to edge applications such as protecting a model's intellectual property and verifying its integrity.

James Marien, Sam Leroux, Bart Dhoedt et al.

Over 60,000 songs are released on Spotify every day, and the competition for the listener's attention is immense. In that regard, the importance of captivating and inviting cover art cannot be underestimated, because it is deeply entangled with a song's character and the artist's identity, and remains one of the most important gateways to lead people to discover music. However, designing cover art is a highly creative, lengthy and sometimes expensive process that can be daunting, especially for non-professional artists. For this reason, we propose a novel deep-learning framework to generate cover art guided by audio features. Inspired by VQGAN-CLIP, our approach is highly flexible because individual components can easily be replaced without the need for any retraining. This paper outlines the architectural details of our models and discusses the optimization challenges that emerge from them. More specifically, we will exploit genetic algorithms to overcome bad local minima and adversarial examples. We find that our framework can generate suitable cover art for most genres, and that the visual features adapt themselves to audio feature changes. Given these results, we believe that our framework paves the road for extensions and more advanced applications in audio-guided visual generation tasks.

Sander De Coninck, Wei-Cheng Wang, Sam Leroux et al.

Camera sensors are increasingly being combined with machine learning to perform various tasks such as intelligent surveillance. Due to its computational complexity, most of these machine learning algorithms are offloaded to the cloud for processing. However, users are increasingly concerned about privacy issues such as function creep and malicious usage by third-party cloud providers. To alleviate this, we propose an edge-based filtering stage that removes privacy-sensitive attributes before the sensor data are transmitted to the cloud. We use state-of-the-art image manipulation techniques that leverage disentangled representations to achieve privacy filtering. We define opt-in and opt-out filter operations and evaluate their effectiveness for filtering private attributes from face images. Additionally, we examine the effect of naturally occurring correlations and residual information on filtering. We find the results promising and believe this elicits further research on how image manipulation can be used for privacy preservation.

Ciem Cornelissen, Sam Leroux, Pieter Simoens

Self-supervised learning has emerged as a powerful paradigm for learning visual representations without manual annotations, yet most methods still operate on a single modality and therefore miss the complementary structure available from heterogeneous sensors. We present Le MuMo JEPA, a self-supervised framework that learns unified representations from RGB images and aligned companion modalities. In our driving experiments, the second modality is camera-aligned LiDAR depth; we also evaluate RGB-thermal training and transfer on the Teledyne FLIR ADAS benchmark. Our approach extends LeJEPA to the multi-modal setting by learning fusion tokens that act as a latent bottleneck between modality-specific patch stems inside a shared transformer. Our default model employs a pruned fusion strategy: after an initial cross-modal attention layer, modality-specific tokens are dropped, forcing cross-modal information into the shared fusion-token grid as an efficient latent bottleneck before Sketched Isotropic Gaussian Regularization (SIGReg) is applied to the joint multimodal CLS embedding. On Waymo, Le MuMo JEPA gives the strongest performance-efficiency trade-off on downstream patch probes among the from-scratch multimodal baselines, improving CenterNet detection and dense depth while remaining competitive on segmentation. Under from-scratch training on nuScenes, Le MuMo JEPA remains the strongest model, and it also gives the best FLIR results, especially after Waymo-initialized fine-tuning. It also retains the best overall accuracy-efficiency balance in our study at substantially lower compute, memory, and estimated training time.

Vishisht Sharma, Sam Leroux, Lisa Landuyt et al.

Effective disaster response relies on rapid disaster response, where oblique aerial video is the primary modality for initial scouting due to its ability to maximize spatial coverage and situational awareness in limited flight time. However, the on-board processing of high-resolution oblique streams is severely bottlenecked by the strict Size, Weight, and Power (SWaP) constraints of Unmanned Aerial Vehicles (UAVs). The computational density required to process these wide-field-of-view streams precludes low-latency inference on standard edge hardware. To address this, we propose Temporal Token Reuse (TTR), an adaptive inference framework capable of accelerating video segmentation on embedded devices. TTR exploits the intrinsic spatiotemporal redundancy of aerial video by formulating image patches as tokens; it utilizes a lightweight similarity metric to dynamically identify static regions and propagate their precomputed deep features, thereby bypassing redundant backbone computations. We validate the framework on standard benchmarks and a newly curated Oblique Floodwater Dataset designed for hydrological monitoring. Experimental results on edge-grade hardware demonstrate that TTR achieves a 30% reduction in inference latency with negligible degradation in segmentation accuracy (< 0.5% mIoU). These findings confirm that TTR effectively shifts the operational Pareto frontier, enabling high-fidelity, real-time oblique video understanding for time-critical remote sensing missions

Sander De Coninck, Emilio Gamba, Bart Van Doninck et al.

The adoption of AI-powered computer vision in industry is often constrained by the need to balance operational utility with worker privacy. Building on our previously proposed privacy-preserving framework, this paper presents its first comprehensive validation on real-world data collected directly by industrial partners in active production environments. We evaluate the framework across three representative use cases: woodworking production monitoring, human-aware AGV navigation, and multi-camera ergonomic risk assessment. The approach employs learned visual transformations that obscure sensitive or task-irrelevant information while retaining features essential for task performance. Through both quantitative evaluation of the privacy-utility trade-off and qualitative feedback from industrial partners, we assess the framework's effectiveness, deployment feasibility, and trust implications. Results demonstrate that task-specific obfuscation enables effective monitoring with reduced privacy risks, establishing the framework's readiness for real-world adoption and providing cross-domain recommendations for responsible, human-centric AI deployment in industry.

Meelis Lootus, Kartik Thakore, Sam Leroux et al.

Although deep neural networks are typically computationally expensive to use, technological advances in both the design of hardware platforms and of neural network architectures, have made it possible to use powerful models on edge devices. To enable widespread adoption of edge based machine learning, we introduce a set of open-source tools that make it easy to deploy, update and monitor machine learning models on a wide variety of edge devices. Our tools bring the concept of containerization to the TinyML world. We propose to package ML and application logic as containers called Runes to deploy onto edge devices. The containerization allows us to target a fragmented Internet-of-Things (IoT) ecosystem by providing a common platform for Runes to run across devices.

Sander De Coninck, Sam Leroux, Pieter Simoens

Mitigating bias in machine learning models is a critical endeavor for ensuring fairness and equity. In this paper, we propose a novel approach to address bias by leveraging pixel image attributions to identify and regularize regions of images containing significant information about bias attributes. Our method utilizes a model-agnostic approach to extract pixel attributions by employing a convolutional neural network (CNN) classifier trained on small image patches. By training the classifier to predict a property of the entire image using only a single patch, we achieve region-based attributions that provide insights into the distribution of important information across the image. We propose utilizing these attributions to introduce targeted noise into datasets with confounding attributes that bias the data, thereby constraining neural networks from learning these biases and emphasizing the primary attributes. Our approach demonstrates its efficacy in enabling the training of unbiased classifiers on heavily biased datasets.

Sander De Coninck, Emilio Gamba, Bart Van Doninck et al.

Musculoskeletal disorders (MSDs) are a leading cause of injury and productivity loss in the manufacturing industry, incurring substantial economic costs. Ergonomic assessments can mitigate these risks by identifying workplace adjustments that improve posture and reduce strain. Camera-based systems offer a non-intrusive, cost-effective method for continuous ergonomic tracking, but they also raise significant privacy concerns. To address this, we propose a privacy-aware ergonomic assessment framework utilizing machine learning techniques. Our approach employs adversarial training to develop a lightweight neural network that obfuscates video data, preserving only the essential information needed for human pose estimation. This obfuscation ensures compatibility with standard pose estimation algorithms, maintaining high accuracy while protecting privacy. The obfuscated video data is transmitted to a central server, where state-of-the-art keypoint detection algorithms extract body landmarks. Using multi-view integration, 3D keypoints are reconstructed and evaluated with the Rapid Entire Body Assessment (REBA) method. Our system provides a secure, effective solution for ergonomic monitoring in industrial environments, addressing both privacy and workplace safety concerns.

Ciem Cornelissen, Sam Leroux, Pieter Simoens

Unmanned Aerial Vehicles (UAVs) combined with Hyperspectral imaging (HSI) offer potential for environmental and agricultural applications by capturing detailed spectral information that enables the prediction of invisible features like biochemical leaf properties. However, the data-intensive nature of HSI poses challenges for remote devices, which have limited computational resources and storage. This paper introduces an Online Hyperspectral Simple Linear Iterative Clustering algorithm (OHSLIC) framework for real-time tree phenotype segmentation. OHSLIC reduces inherent noise and computational demands through adaptive incremental clustering and a lightweight neural network, which phenotypes trees using leaf contents such as chlorophyll, carotenoids, and anthocyanins. A hyperspectral dataset is created using a custom simulator that incorporates realistic leaf parameters, and light interactions. Results demonstrate that OHSLIC achieves superior regression accuracy and segmentation performance compared to pixel- or window-based methods while significantly reducing inference time. The method`s adaptive clustering enables dynamic trade-offs between computational efficiency and accuracy, paving the way for scalable edge-device deployment in HSI applications.

Mattijs Baert, Sam Leroux, Pieter Simoens

Learning from Demonstrations (LfD) and Reinforcement Learning (RL) have enabled robot agents to accomplish complex tasks. Reward Machines (RMs) enhance RL's capability to train policies over extended time horizons by structuring high-level task information. In this work, we introduce a novel LfD approach for learning RMs directly from visual demonstrations of robotic manipulation tasks. Unlike previous methods, our approach requires no predefined propositions or prior knowledge of the underlying sparse reward signals. Instead, it jointly learns the RM structure and identifies key high-level events that drive transitions between RM states. We validate our method on vision-based manipulation tasks, showing that the inferred RM accurately captures task structure and enables an RL agent to effectively learn an optimal policy.

Mattijs Baert, Sam Leroux, Pieter Simoens

The alignment of autonomous agents with human values is a pivotal challenge when deploying these agents within physical environments, where safety is an important concern. However, defining the agent's objective as a reward and/or cost function is inherently complex and prone to human errors. In response to this challenge, we present a novel approach that leverages one-class decision trees to facilitate learning from expert demonstrations. These decision trees provide a foundation for representing a set of constraints pertinent to the given environment as a logical formula in disjunctive normal form. The learned constraints are subsequently employed within an oracle constrained reinforcement learning framework, enabling the acquisition of a safe policy. In contrast to other methods, our approach offers an interpretable representation of the constraints, a vital feature in safety-critical environments. To validate the effectiveness of our proposed method, we conduct experiments in synthetic benchmark domains and a realistic driving environment.

Ciem Cornelissen, Sander De Coninck, Axel Willekens et al.

This paper presents an end-to-end, IoT-enabled robotic system for the non-destructive, real-time, and spatially-resolved mapping of grape yield and quality (Brix, Acidity) in vineyards. The system features a comprehensive analytical pipeline that integrates two key modules: a high-performance model for grape bunch detection and weight estimation, and a novel deep learning framework for quality assessment from hyperspectral (HSI) data. A critical barrier to in-field HSI is the ``domain shift" caused by variable illumination. To overcome this, our quality assessment is powered by the Light-Invariant Spectral Autoencoder (LISA), a domain-adversarial framework that learns illumination-invariant features from uncalibrated data. We validated the system's robustness on a purpose-built HSI dataset spanning three distinct illumination domains: controlled artificial lighting (lab), and variable natural sunlight captured in the morning and afternoon. Results show the complete pipeline achieves a recall (0.82) for bunch detection and a $R^2$ (0.76) for weight prediction, while the LISA module improves quality prediction generalization by over 20% compared to the baselines. By combining these robust modules, the system successfully generates high-resolution, georeferenced data of both grape yield and quality, providing actionable, data-driven insights for precision viticulture.

Matteo Cardoni, Sam Leroux

As deep neural networks are increasingly deployed in dynamic, real-world environments, relying on a single static model is often insufficient. Changes in input data distributions caused by sensor drift or lighting variations necessitate continual model adaptation. In this paper, we propose a hybrid training methodology that enables efficient on-device domain adaptation by combining the strengths of Backpropagation and Predictive Coding. The method begins with a deep neural network trained offline using Backpropagation to achieve high initial performance. Subsequently, Predictive Coding is employed for online adaptation, allowing the model to recover accuracy lost due to shifts in the input data distribution. This approach leverages the robustness of Backpropagation for initial representation learning and the computational efficiency of Predictive Coding for continual learning, making it particularly well-suited for resource-constrained edge devices or future neuromorphic accelerators. Experimental results on the MNIST and CIFAR-10 datasets demonstrate that this hybrid strategy enables effective adaptation with a reduced computational overhead, offering a promising solution for maintaining model performance in dynamic environments.

Mattijs Baert, Pietro Mazzaglia, Sam Leroux et al.

When deploying artificial agents in real-world environments where they interact with humans, it is crucial that their behavior is aligned with the values, social norms or other requirements of that environment. However, many environments have implicit constraints that are difficult to specify and transfer to a learning agent. To address this challenge, we propose a novel method that utilizes the principle of maximum causal entropy to learn constraints and an optimal policy that adheres to these constraints, using demonstrations of agents that abide by the constraints. We prove convergence in a tabular setting and provide an approximation which scales to complex environments. We evaluate the effectiveness of the learned policy by assessing the reward received and the number of constraint violations, and we evaluate the learned cost function based on its transferability to other agents. Our method has been shown to outperform state-of-the-art approaches across a variety of tasks and environments, and it is able to handle problems with stochastic dynamics and a continuous state-action space.

Sander De Coninck, Sam Leroux, Pieter Simoens

Crowd management relies on inspection of surveillance video either by operators or by object detection models. These models are large, making it difficult to deploy them on resource constrained edge hardware. Instead, the computations are often offloaded to a (third party) cloud platform. While crowd management may be a legitimate application, transferring video from the camera to remote infrastructure may open the door for extracting additional information that are infringements of privacy, like person tracking or face recognition. In this paper, we use adversarial training to obtain a lightweight obfuscator that transforms video frames to only retain the necessary information for person detection. Importantly, the obfuscated data can be processed by publicly available object detectors without retraining and without significant loss of accuracy.

Mattijs Baert, Sam Leroux, Pieter Simoens

The widespread deployment of surveillance cameras for facial recognition gives rise to many privacy concerns. This study proposes a privacy-friendly alternative to large scale facial recognition. While there are multiple techniques to preserve privacy, our work is based on the minimization principle which implies minimizing the amount of collected personal data. Instead of running facial recognition software on all video data, we propose to automatically extract a high quality snapshot of each detected person without revealing his or her identity. This snapshot is then encrypted and access is only granted after legal authorization. We introduce a novel unsupervised face image quality assessment method which is used to select the high quality snapshots. For this, we train a variational autoencoder on high quality face images from a publicly available dataset and use the reconstruction probability as a metric to estimate the quality of each face crop. We experimentally confirm that the reconstruction probability can be used as biometric quality predictor. Unlike most previous studies, we do not rely on a manually defined face quality metric as everything is learned from data. Our face quality assessment method outperforms supervised, unsupervised and general image quality assessment methods on the task of improving face verification performance by rejecting low quality images. The effectiveness of the whole system is validated qualitatively on still images and videos.

Bo Li, Sam Leroux, Pieter Simoens

Automating the analysis of surveillance video footage is of great interest when urban environments or industrial sites are monitored by a large number of cameras. As anomalies are often context-specific, it is hard to predefine events of interest and collect labelled training data. A purely unsupervised approach for automated anomaly detection is much more suitable. For every camera, a separate algorithm could then be deployed that learns over time a baseline model of appearance and motion related features of the objects within the camera viewport. Anything that deviates from this baseline is flagged as an anomaly for further analysis downstream. We propose a new neural network architecture that learns the normal behavior in a purely unsupervised fashion. In contrast to previous work, we use latent code predictions as our anomaly metric. We show that this outperforms reconstruction-based and frame prediction-based methods on different benchmark datasets both in terms of accuracy and robustness against changing lighting and weather conditions. By decoupling an appearance and a motion model, our model can also process 16 to 45 times more frames per second than related approaches which makes our model suitable for deploying on the camera itself or on other edge devices.

Xander Steenbrugge, Sam Leroux, Tim Verbelen et al.

In this work we explore the generalization characteristics of unsupervised representation learning by leveraging disentangled VAE's to learn a useful latent space on a set of relational reasoning problems derived from Raven Progressive Matrices. We show that the latent representations, learned by unsupervised training using the right objective function, significantly outperform the same architectures trained with purely supervised learning, especially when it comes to generalization.

Sam Leroux, Tim Verbelen, Pieter Simoens et al.

Deep neural networks require large amounts of resources which makes them hard to use on resource constrained devices such as Internet-of-things devices. Offloading the computations to the cloud can circumvent these constraints but introduces a privacy risk since the operator of the cloud is not necessarily trustworthy. We propose a technique that obfuscates the data before sending it to the remote computation node. The obfuscated data is unintelligible for a human eavesdropper but can still be classified with a high accuracy by a neural network trained on unobfuscated images.

Sam Leroux, Pavlo Molchanov, Pieter Simoens et al.

Deep residual networks (ResNets) made a recent breakthrough in deep learning. The core idea of ResNets is to have shortcut connections between layers that allow the network to be much deeper while still being easy to optimize avoiding vanishing gradients. These shortcut connections have interesting side-effects that make ResNets behave differently from other typical network architectures. In this work we use these properties to design a network based on a ResNet but with parameter sharing and with adaptive computation time. The resulting network is much smaller than the original network and can adapt the computational cost to the complexity of the input image.

Sam Leroux, Steven Bohez, Tim Verbelen et al.

Previous work has shown that it is possible to train deep neural networks with low precision weights and activations. In the extreme case it is even possible to constrain the network to binary values. The costly floating point multiplications are then reduced to fast logical operations. High end smart phones such as Google's Pixel 2 and Apple's iPhone X are already equipped with specialised hardware for image processing and it is very likely that other future consumer hardware will also have dedicated accelerators for deep neural networks. Binary neural networks are attractive in this case because the logical operations are very fast and efficient when implemented in hardware. We propose a transfer learning based architecture where we first train a binary network on Imagenet and then retrain part of the network for different tasks while keeping most of the network fixed. The fixed binary part could be implemented in a hardware accelerator while the last layers of the network are evaluated in software. We show that a single binary neural network trained on the Imagenet dataset can indeed be used as a feature extractor for other datasets.

Pieter Van Molle, Tim Verbelen, Steven Bohez et al.

Reinforcement learning is a proven technique for an agent to learn a task. However, when learning a task using reinforcement learning, the agent cannot distinguish the characteristics of the environment from those of the task. This makes it harder to transfer skills between tasks in the same environment. Furthermore, this does not reduce risk when training for a new task. In this paper, we introduce an approach to decouple the environment characteristics from the task-specific ones, allowing an agent to develop a sense of survival. We evaluate our approach in an environment where an agent must learn a sequence of collection tasks, and show that decoupled learning allows for a safer utilization of prior knowledge.

Sam Leroux, Steven Bohez, Cedric De Boom et al.

In this paper we propose a technique which avoids the evaluation of certain convolutional filters in a deep neural network. This allows to trade-off the accuracy of a deep neural network with the computational and memory requirements. This is especially important on a constrained device unable to hold all the weights of the network in memory.

Cedric De Boom, Sam Leroux, Steven Bohez et al.

We present four training and prediction schedules from the same character-level recurrent neural network. The efficiency of these schedules is tested in terms of model effectiveness as a function of training time and amount of training data seen. We show that the choice of training and prediction schedule potentially has a considerable impact on the prediction effectiveness for a given training budget.