As a part of the European project ArtiSaneFood, the primary goal of this collaborative work between ANSES, CNIEL and L2S is to establish efficient bio-intervention strategies for cheese producers in France, to “economically” reduce the risk of Haemolytic Uremic Syndrome (HUS) caused by Shiga-Toxin producing Escherichia Coli (STEC) present in raw-milk soft cheese. This translates into a multiobjective optimization problem of a stochastic simulator based on a quantitative risk assessment (QRA) model proposed by Perrin et al. (2014), to estimate the Pareto optimal solutions for the process intervention parameters. 

Popular benchmarks in Computer Vision suffer from a Western-centric bias that leads to a distribution shift problem when trying to deploy Machine Learning systems in developing countries. Palliating this problem using the same data generation methods in poorly represented countries will likely bring the same bias that were initially observed. In this paper, we propose an adaptation of the MS COCO data generation methodology that address this issue and show how the web scraping methods nests geographical distribution shifts.

Although there are over a hundred different sets of AI ethics principles and guidelines to steer the ethical (trustworthy, responsible) design and implementation of the technology, little has been said in these guidelines about the sustainability of AI/ML. The notion ‘sustainability’ has risen to extraordinary relevance in the face of the current climate crisis and researchers have since picked up on the sustainability impacts of AI. My main research objective is to explore the conceptual foundations of ‘sustainability’ in order to devise a normative sustainability framework that will provide a fruitful addition to this young and ongoing debate in AI ethics.

The last decade has seen a growth in the number of attacks with severe economic and privacy damages, which reveals the need for Network Intrusion Detection Systems.  In this work we propose a flow-based solution for network cyber-attacks detection in the edge context. The framework takes as input a flow-based graph structure and outputs a maliciousness score using graph neural network (GNN) algorithms to classify flows into normal and malicious.

Object State Detection (SD) is a challenging task of great importance to fields, such as Computer Vision, Robotics, among others. In this work, we present results
investigating on how the incorporation of Common-Sense Knowledge in the form of Knowledge Graphs can improve the overall performance on SD.

Deep Learning models can be extremely complex to configure to perform well. Neural Architecture Search (NAS) offers to navigate these spaces automatically, but NAS remains costly in terms of computational resources. Our goal is to generate architectures achieving trade-offs between performance and sustainability by exploiting network topologies and domain constraints.

Various Technologies depend on neural network decisions. However, guaranteeing their behaviour is still a significant challenge. To this end, we are keen to co-develop an easy-to-learn and verify NN architecture. Our focus is on polynomial NN with quadratic activation functions in which formal verification remains an undiscovered area for this type of network where the most excellent potential might be found.

In this work, we are going to present a new deep framework, allowing us to merge multi-MRI image segmentation results using the formalism of Dempster-Shafer theory and contextual discounting while considering the reliability of different modalities relative to different classes..

A main issue of Machine Learning models is knowing when their predictions can be trusted. Here we explore how to express the operational domain (or “domain of validity”) of a classifier, using generative models to learn powerful data representations and generate extreme examples. First results show that generative models can be used to find the visual attributes that impact the most the accuracy of a given classifier.

Text Sensors are installed on every Air Liquide plant asset to produce securely, monitor and optimize the plant performance. In contrast to process problems, data anomalies are problems that are not fully handled nowadays, which can have a significant impact on the plant operations. Since no labels are identified for these types of anomalies, the proposed solution is to perform active learning with expert-in-the-loop to iteratively detect and validate novel data anomalies.

In an unstable and critical time marked by the COVID-19 pandemic, the Russian war against Ukraine, and the climate crisis, imminent changes and constraints are inevitable. In this research project, we aim to address the issue of how the current global crises influence the state of mind of entrepreneurs who have just started their business (novices) and those who have already founded a business (experts). Based on the collected data, annotated transcriptions are used to identify the mentioned emotions according to Plutchik’s eight basic emotions, to determine their intensity and underlying causes, and to distinguish whether these emotions are positive or negative.

Comparison of boxing performance metrics evaluation between high quality video and compressed and sparse video computer vision analysis. A substantial gain in terms of data storage and computational time for a minimal loss of information is highlighted.

While much researched, anomaly-based network intrusion detection systems (NIDS) have few real-world applications due to a lack of adequate modern datasets and confidence in the often-obscure machine learning (ML) models. Using XAI methods, we show that imbalances of attack types bias the classification of attack classes and SHAP explanations in the NSL-KDD dataset. Analysing attack types directly improves performance, also enabling SHAP explanations that have a better fit with domain expertise. In conclusion, XAI methods like SHAP can be employed to find and debug issues with NIDS data and machine learning models..

We quantify the robustness of a trained network to input uncertainties with a stochastic simulation inspired by the field of Statistical Reliability Engineering. The robustness assessment is cast as a statistical hypothesis test: the network is deemed as locally robust if the estimated probability of failure is lower than a critical level. .

The growing implementation of ML-based Diagnostic Decision Support Systems (DDSSs) in various medical fields pose new challenges to the care-relationship between patients and medical experts. In the context of DDSSs especially the connection and tension between transparency and trust plays a crucial role. Since DDSSs do not work in isolation but within other systems including human agents in different roles, e.g., patients, doctors, nurses, and system developers, it is crucial to ask for whom these systems need to be transparent as well as whom or what is trusted or relied upon. The main question here then is, to what extent care-relationships are altered, if there is a non-fully transparent AI mechanism inserted in the diagnostic process and what implications that has for interpersonal and institutional/public trust in Medical Expertise.

Different works have shown that GANs implicitly encode the semantics of the training data in the latent space. It is thus possible to exploit the existing structure to extract various controls over the generated data. We will present two applications. First, an application to control binary facial attributes (e.g., glasses) together with a method to address the problem of entanglement between attributes. Second, an extension of existing methods to control a different type of attribute: the number of objects in a scene.

Centralized Federated Learning, Gossip and Blockchain-enabled federated learning are some of the most appealing solutions to train a machine learning model over distributed data-sources. We propose a comparison between these techniques combining standard performance indicators, like accuracy, with indicators to quantify the efficiency of these algorithms. 

Access to common sense knowledge is receiving renewed interest for developing neuro-symbolic AI systems or debugging deep learning models. Little is currently understood about the types of knowledge that can be gathered using existing knowledge elicitation methods. Moreover, these methods fall short of meeting the evolving requirements of several downstream AI tasks. To this end, collecting broad and tacit knowledge, in addition to negative or discriminative knowledge can be highly useful. Addressing this research gap, we developed a novel game with a purpose, ‘FindItOut’, to elicit different types of knowledge from human players through easily configurable game mechanics.

Our work presents the modelling of cellular network datasets, namely Charging Data Records (CDRs), that fulfil real-world CDRs attributes. We reproduce realistic spatio-temporal human mobility behaviour and leverage Long-Short Term Memory to model individual nodes’ traffic behaviour. Based on a lifelike structure of the cellular network architecture and social life, we combine these models into complete and realistic CDRs that reproduce the daily cellular urban life and are highly valuable.

Leakages in pressured air systems are a frequent and difficult to trace source of energy losses. Detecting them requires a repeated time-consuming manual search, during which a worker usually needs to scan the whole pipe duct and mark any leakages by hand. The leakages can be detected by monitoring the sound pressure level to identify distinctive patterns within a narrow frequency band using an ultrasonic microphone. Typically, attention is paid to exceeding a certain threshold within these patterns, but matching algorithms or machine learning could also be employed to improve the data-based detection of leakages. In the scope of the research project KI4ETA, we attempt to automate this process with a mobile robot efficiently using the same sensing method. While past automation approaches focus on a full scan of the environment, we aim at conducting a targeted search and therefore minimize the necessary time for the detection of leakages.

Our approach leverages visual SLAM technology to get a more information-rich model of its environment. This enables a more targeted search and better plannability to reach good detection positions. It also features the integration of human expert knowledge, to constrict the search to regions of increased probability of leakages. With the use of AI-based image-processing, the environment scan can be limited to relevant directions of the field of view.

This combination of these techniques enables targeted strategies toward efficient and automated leakage detection. It also brings inherent digitization of the search and documentation process. To fully leverage this potential, the whole application is highly integrable with other systems using standard interfaces like OPC-UA for data transfer.

Like in many other domains, machine learning methods have had a tremendous impact on wind energy research in recent years. In wind turbine power curve modelling, for example, deep ANNs have shown significant advantages over conventional methods. Despite their impressive predictive abilities, however, ANNs are often criticized as opaque black-boxes with no physical understanding of the system they describe. We apply Shapley Values to uncover the strategies learned by ANNs trained to model turbine output based on operational data. The results have significant implications for model selection and wind turbine monitoring applications.

Data to analyse for attacks detection in network are very unbalanced in general. This fact makes it hard for statistical anomalies to prove themselves useful in attacks detection.

Moreover, there is a need to be able to analyse the relation between the different elements in the network and answer to the need of new features with less statistical bias.

The graph-based approach seemed like a meaningful choice, since this approach can show a representation of the relation in the network.

Graph metrics can then be used as both representations of temporal and spatial behaviours in the network, especially community metrics and their evolution in a dynamic graph.

We propose an empirical study to see whether learning with convolutional neural networks (CNNs) on entropy coded data is possible.

But entropy coding destroys the properties needed for a CNN to work well. Thus, learning on such coded data with a CNN should prove impossible. Our experimental results show otherwise: learning in such difficult conditions is still possible albeit with reduced accuracy.

In the context of neurofeedback (NF) using simultaneous acquisitions with electroencephalography (EEG) and functional magnetic resonance imaging (fMRI), our goal is to enhance EEG-based NF using fMRI information. In the future, this will help to reduce the use of the heavy and time-consuming MRI modality. To do so, we propose a method based on a recurrent neural network (RNN) that consists in learning a model from simultaneous EEG-fMRI acquisitions to predict NF-fMRI scores with EEG signals features alone. 

In the last decades, logical models have emerged as a successful paradigm for capturing and predicting the behaviour of systems of molecular interactions. Intuitively, they consist in sampling the abundance of each kind of biochemical entity within finite sets of intervals and deriving transitions accordingly. Whereas formally sound derivation from more precise descriptions (such as from reaction networks) includes many fictitious behaviours, direct modelling usually favours dominant interactions with no guarantee on the behaviours that are neglected.

In this work, we formalize a sound and accurate discretization method for behaviours emerging from stochastic reaction networks. To do so, we design overlapping intervals to introduce a minimal effort for the system to go back to an abstract region of states, hence limiting fictitious oscillations in the derived models. Then, we compute for pairs of transitions (in the derived model) bounds on the probabilities on which one will occur first. We illustrate our ideas on two case studies and demonstrate how techniques from Abstract Interpretation can be used to construct more precise discretization methods, while providing a framework to further investigate the underlying structure of logical models.

The poster presents a baseline approach for user identification based on multimodal behavioral biometrics using classical machine learning and deep learning. Two behavioral biometrics have been used : human activities captured by a smartphone and keystroke dynamics on a laptop. A generic workflow has been used for the classical approach using Orange data mining software. Recent successful deep learning approaches for time series classification on python have been used.

Flexible production in modern smart factories requires effective support of human workers and smooth cooperation with assisting robots. For this very reason, we base any assistive functionality on high-level intentions. In the InCoRAP project, recognition of high-level intentions is based on gaze estimation and object detection (OD) using probabilistic process models. For optimizing the object detection training process to learn consecutive tasks without losing performance on the previous ones, we investigate different continual object detection approaches. We can reduce the training time and memory size and hence power consumption by utilizing the continual object detection approaches for updating the previous model instead of training the novice model which leads to the design of systems binding with sustainable AI principles.

The thesis will focus on the development of a decision support agent for the operation of the electricity transmission grid through the decentralized multi-agent reinforcement learning (MARL) approach. We will consider the constraints of each sub-agent, while trying to minimize changes and uncertainties associated with unpredictability/instability of the production of green electricity sources as well as related to incomplete information sources such as weather forecasts. This subject stems from the energy transition and the smooth development and integration of renewable energies. 

In real-world ML problems, models tend to suffer some degradation in terms of performance over time, due to changes in the concept of what is first learned by the model (concept drift), or due to significant changes in the variables used by the model (data/covariate drift). We present Frouros, a Python library capable of detecting drift using both supervised and unsupervised drift detection methods, and which can be easily integrated with scikit-learn.

Urinary tract infections (UTI) belong to the most common clinically relevant bacterial infections. 1 in 3 women worldwide will have at least one UTI by 24 years of age and 40 – 50% of women will experience one UTI during their lifetime with 44% experiencing recurrences. In this project, using a clinical dataset of brightfield microscopy of patient’s urine with few annotated samples, we aim to develop a diagnostic phenotype quantification workflow using label-efficient machine learning (ML) approaches. There are several challenges to the clinical dataset at hand. Firstly, in the absence of specific labelling for phenotype-relevant objects in the micrographs ground truth is ambiguous. Secondly, obtaining manual annotations is laborious and requires highly-skilled annotators. Thirdly, the variation in scale and shape of a particular type of phenotype-relevant objects is challenging for the instance segmentation. To address these, first we develop a deep learning (DL) model for object detection and binary segmentation in clinical samples of patients. Since the pixel-level microscopy annotation is time-consuming and requires expert knowledge we explore label-efficient weakly or self-supervised approaches as pretext tasks to pre-train our DL model. Furthermore, to use the full extent of the optical resolution of the brightfield microscopy, as well as employ data augmentation and class balancing we use a custom generator of micrograph patches. Next, to obtain weak multi-class annotations for objects present in the micrographs we employ feature extraction with subsequent K-Means clustering. Finally, we train or fine-tune our DL model end-to-end and provide an evaluation of label-efficient techniques. 

Multimorbidities, the presence of multiple diseases or conditions in a patient, are strongly dependent on age, and they change with patients’ aging. We propose a novel approach on modelling dynamical comorbidity networks from longitudinal population-wide healthcare data from Austria. We aim to understand disease trajectories and in which directions they change overage.

Several embedding methods were proposed to learn low dimensional vector representations of entities and relations of a KG. Such representations facilitate the link prediction task in the service of inference and KG completion. The power of embedding methods is often criticized for their lack of explainability. In this context, it is important to explain link predictions. As for explainability, it constitutes a thriving research question, especially when it comes to analysing KGs with their rich semantics rooted in description logics. In this study, a new rule mining method is developed based on learned embeddings. The extracted rules are used as support for explainable link prediction. Check my recent work (https://doi.org/10.1016/j.knosys.2022.109083) for details.

Traffic forecasting is largely covered in the literature, in the last year with the novel Deep Learning (DL) solutions. However, the sustainability of this approach is still an open issue that we want to investigate in this work. Different learning solutions will be analysed, with the aim of decreasing the energy consumption in the training phase of the DL models and maintaining the accuracy. We proposed a distributed solution to forecast traffic flow based on continual learning paradigm, and we compared it with a standard centralized model and a local approach, in which models are trained ad hoc for each sensor of the dataset. 

Pre-trained models are increasingly used for transfer learning. Often it is unclear what data the models have been trained on and what biases (e.g., spurious correlations in the data) they have been subjected to. We propose a soft-pruning strategy using the little data usually available for down-stream tasks to pre-emptively robustify any given model against using spurious correlations learned during pretraining. 

In long-running data streams concept drift denotes the phenomenon that the prior or posterior probabilities—P(X) or P(Y|X)—governing the data stream may change such that other models deployed on the stream must be adapted. Sensor streams observing marine environments combine several challenges not addressed by state-of-the-art concept drift detectors in their entirety, including but not limited to fully unlabelled data, seasonality, sensor outages and drift. Based on this analysis, possible approaches for a dedicated concept drift detector for marine sensor streams are outlined.

Federated Learning (FL) allows multiple devices to learn a shared model without exchanging private data. A typical scenario involves using constrained devices in a massively distributed environment combining Cloud, Fog and Edge resources, also called Computing Continuum (CC).

Running FL workflows across the CC involves frequent deployment and monitoring in large scale and heterogeneous environments while considering several objectives such as privacy preservation, quality of the prediction and resource consumption.

To this purpose, additional tools must be used to adapt ML workflows and better support deployment of FL Systems. We propose a framework to automatically deploy FL workloads in heterogeneous environments using formal description of the underlying infrastructure, hyperparameter optimization and monitoring tools to ease management of the system.

The unstoppable improvements in data analysis techniques for knowledge extraction and decision-making lead to the evolution of techniques for the secure publication and sharing of data. With this aim, the implementation of pyCANON is presented. pyCANON is a Python library that can be used to know the level of anonymity of a dataset (and thus publish or share it while being aware of the risks involved). Nine different anonymization techniques will be used for this purpose.

Vector Symbolic Architectures (VSAs) combine a high-dimensional vector space with a set of carefully designed operators to perform symbolic computations with large numerical vectors. Major goals are the exploitation of their representational power and ability to deal with fuzziness and ambiguity. The presentation explains our experience in applying VSAs in computer vision and signal processing, specifically visual place recognition and time series classification. In visual place recognition, we can enrich the descriptor vector of an image with additional information, such as spatial semantic information, without increasing the resulting vector representation. In time series classification, we integrated the principles of a VSA into a state-of-the-art time series classification algorithm to provide explicit global time encoding⁠. This prevents the original method from failing in special cases where global context is important to distinguish signals. 

This poster will present how automatic speech processing can help better understand the representation of women and men in broadcast TV and radio and highlight its evolutions.

Interest in research in laser plasma acceleration (LPA) processes is increasing year by year and reveals wide range of its applications (e.g. tumour diagnoses and therapy or computer chips manufacturing). Optimization of control over the particle acceleration phenomena requires complex technical equipment and comprehensive experimental research that must be supported by theoretical findings and numerical simulations. To connect experimental and theoretical results, process them to provide strategies for stabilizing and improvement of particle acceleration we suggest a deep learning based surrogate model. In our work we propose a generative (normalizing flows based) surrogate model that can describe an electron bunch transformation in the simulated beam transport – a sequence of optical elements that amplifies desired properties of an accelerated bunch of electrons. In future such kind of model can provide more insights on its internal dynamics, that cannot be observed during the experiment.

This work proposes a solution based on GAN to translate genuine fingerprint images into what they would look like if they were created from known spoof (fake fingerprint) materials.

This is an application of style transfer of GAN which can be used for certification of biometric systems.