Développer un programme capable de jouer à n’importe quel jeu de stratégie, souvent désigné par le General Game Playing (GGP) constitue un des Graal de l’intelligence artificielle. Les compétitions GGP, où chaque jeu est représenté par un ensemble de règles logiques au travers du Game Description Language (GDL), ont conduit la recherche à confronter de nombreuses approches incluant les méthodes de type Monte Carlo, la construction automatique de fonctions d’évaluation, ou la programmation logique et ASP. De par cette thèse, nous proposons une nouvelle approche dirigée par les contraintes stochastiques.Dans un premier temps, nous nous concentrons sur l’élaboration d’une traduction de GDL en réseauxde contraintes stochastiques (SCSP) dans le but de fournir une représentation dense des jeux de stratégies et permettre la modélisation de stratégies.Par la suite, nous exploitons un fragment de SCSP au travers d’un algorithme dénommé MAC-UCBcombinant l’algorithme MAC (Maintaining Arc Consistency) utilisé pour résoudre chaque niveau duSCSP tour après tour, et à l’aide de UCB (Upper Confidence Bound) afin d’estimer l’utilité de chaquestratégie obtenue par le dernier niveau de chaque séquence. L’efficacité de cette nouvelle technique sur les autres approches GGP est confirmée par WoodStock, implémentant MAC-UCB, le leader actuel du tournoi continu de GGP.Finalement, dans une dernière partie, nous proposons une approche alternative à la détection de symétries dans les jeux stochastiques, inspirée de la programmation par contraintes. Nous montrons expérimentalement que cette approche couplée à MAC-UCB, surpasse les meilleures approches du domaine et a permis à WoodStock de devenir champion GGP 2016. / The ability for a computer program to effectively play any strategic game, often referred to General Game Playing (GGP), is a key challenge in AI. The GGP competitions, where any game is represented according to a set of logical rules in the Game Description Language (GDL), have led researches to compare various approaches, including Monte Carlo methods, automatic constructions of evaluation functions, logic programming, and answer set programming through some general game players. In this thesis, we offer a new approach driven by stochastic constraints. We first focus on a translation process from GDL to stochastic constraint networks (SCSP) in order to provide compact representations of strategic games and to model strategies. In a second part, we exploit a fragment of SCSP through an algorithm called MAC-UCB by coupling the MAC (Maintaining Arc Consistency) algorithm, used to solve each stage of the SCSP in turn, together with the UCB (Upper Confidence Bound) policy for approximating the values of those strategies obtained by the last stage in the sequence. The efficiency of this technical on the others GGP approaches is confirmed by WoodStock, implementing MAC-UCB, the actual leader on the GGP Continuous Tournament. Finally, in the last part, we propose an alternative approach to symmetry detection in stochastic games, inspired from constraint programming techniques. We demonstrate experimentally that MAC-UCB, coupled with our constranit-based symmetry detection approach, significantly outperforms the best approaches and made WoodStock the GGP champion 2016.

Le changement de croyances vise à trouver des moyens adéquats pour faire évoluer les croyances d'un agent lorsqu'il est confronté à de nouvelles informations. Dans la plupart des travaux sur la révision de croyances, l'ensemble de croyances d'un agent est composé de croyances au sujet de l'environnement (le monde) et est représenté par un ensemble de formules de la logique classique. Dans de nombreuses applications, un agent n'est pas seul dans l'environnement, mais le partage avec d'autres agents, qui ont aussi des croyances. Ainsi les croyances sur les croyances des autres agents constituent un élément d'information important pour l'agent, afin d'être en mesure de prendre les meilleures décisions et d'effectuer les meilleures actions. L'utilisation de croyances sur les croyances des autres agents est par exemple cruciale dans la théorie des jeux. Dans cette thèse, nous étudions dans un premier temps les opérateurs de contraction propositionnelle correspondant aux opérateurs de révision de Katsuno et Mendelzon. Nous étudions ensuite une connexion entre les logiques épistémiques et la théorie du changement de croyances, proche de l'approche AGM. Nous nous sommes intéressés à l'utilisation des opérateurs qui modifient les croyances des agents dans les modèles KD45n standard. Cette tâche est plus compliquée que dans le cadre AGM standard, car, dans un contexte multi-agents, les nouvelles informations peuvent prendre différentes formes. Par exemple, chaque nouvelle information peut être observée/transmise/disponible à tous les agents ou seulement à certains d’entre eux. / Belief change is about finding appropriate ways to evolve an agent's beliefs when confronted with new pieces of information. In most works on belief revision, the set of beliefs of an agent is composed of beliefs about the environment (the world) and is represented by a set of formulas of classical logic. In many applications, an agent is not alone in the environment, but sharing with other agents, which also have beliefs. Thus beliefs about the beliefs of other agents are an important piece of information for the agent in order to be able to make the best decisions and perform the best actions. The use of beliefs about the beliefs of other agents is, for exampel, crucial in game theory. In this thesis, we first study the operators of propositional contraction corresponding to the revision operators proposed by Katsuno and Mendelzon. Then, we study a connection between epistemic logics and belief change theory, close to the AGM approach. We are interested in the use of operators that modify agent beliefs in standard KD45n models. This task is more complicated than in the standard AGM framework because, in a multi-agent context, new information can take different forms. For example, each new information can be observed/transmitted/available to all agents or only some of them.

In this work a coherent survey of problems connected with relational knowledge representation and methods for achieving relational knowledge representation were presented. Proposed approach was shown on three applications: economic case, biomedical case and benchmark dataset. All crucial definitions were formulated and three main methods for relation identification problem were shown. Moreover, for specific relational models and observations’ types different identification methods were presented. / Double Diploma Programme, polish supervisor: prof. Jerzy Świątek, Wrocław University of Technology

Abstract Knowledge discovery (KD) technology is used to extract knowledge from large quantities of digital data in an automated fashion. The established process model represents the KD process in a linear and technology-centered manner, as a sequence of transformations that refine raw data into more and more abstract and distilled representations. Any actual KD process, however, has aspects that are not adequately covered by this model. In particular, some of the most important actors in the process are not technological but human, and the operations associated with these actors are interactive rather than sequential in nature. This thesis proposes an augmentation of the established model that addresses this neglected dimension of the KD process. The proposed process model is composed of three sub-models: a data model, a workflow model, and an architectural model. Each sub-model views the KD process from a different angle: the data model examines the process from the perspective of different states of data and transformations that convert data from one state to another, the workflow model describes the actors of the process and the interactions between them, and the architectural model guides the design of software for the execution of the process. For each of the sub-models, the thesis first defines a set of requirements, then presents the solution designed to satisfy the requirements, and finally, re-examines the requirements to show how they are accounted for by the solution. The principal contribution of the thesis is a broader perspective on the KD process than what is currently the mainstream view. The augmented KD process model proposed by the thesis makes use of the established model, but expands it by gathering data management and knowledge representation, KD workflow and software architecture under a single unified model. Furthermore, the proposed model considers issues that are usually either overlooked or treated as separate from the KD process, such as the philosophical aspect of KD. The thesis also discusses a number of technical solutions to individual sub-problems of the KD process, including two software frameworks and four case-study applications that serve as concrete implementations and illustrations of several key features of the proposed process model. / Tiivistelmä Tiedonlouhintateknologialla etsitään automoidusti tietoa suurista määristä digitaalista dataa. Vakiintunut prosessimalli kuvaa tiedonlouhintaprosessia lineaarisesti ja teknologiakeskeisesti sarjana muunnoksia, jotka jalostavat raakadataa yhä abstraktimpiin ja tiivistetympiin esitysmuotoihin. Todellisissa tiedonlouhintaprosesseissa on kuitenkin aina osa-alueita, joita tällainen malli ei kata riittävän hyvin. Erityisesti on huomattava, että eräät prosessin tärkeimmistä toimijoista ovat ihmisiä, eivät teknologiaa, ja että heidän toimintansa prosessissa on luonteeltaan vuorovaikutteista eikä sarjallista. Tässä väitöskirjassa ehdotetaan vakiintuneen mallin täydentämistä siten, että tämä tiedonlouhintaprosessin laiminlyöty ulottuvuus otetaan huomioon. Ehdotettu prosessimalli koostuu kolmesta osamallista, jotka ovat tietomalli, työnkulkumalli ja arkkitehtuurimalli. Kukin osamalli tarkastelee tiedonlouhintaprosessia eri näkökulmasta: tietomallin näkökulma käsittää tiedon eri olomuodot sekä muunnokset olomuotojen välillä, työnkulkumalli kuvaa prosessin toimijat sekä niiden väliset vuorovaikutukset, ja arkkitehtuurimalli ohjaa prosessin suorittamista tukevien ohjelmistojen suunnittelua. Väitöskirjassa määritellään aluksi kullekin osamallille joukko vaatimuksia, minkä jälkeen esitetään vaatimusten täyttämiseksi suunniteltu ratkaisu. Lopuksi palataan tarkastelemaan vaatimuksia ja osoitetaan, kuinka ne on otettu ratkaisussa huomioon. Väitöskirjan pääasiallinen kontribuutio on se, että se avaa tiedonlouhintaprosessiin valtavirran käsityksiä laajemman tarkastelukulman. Väitöskirjan sisältämä täydennetty prosessimalli hyödyntää vakiintunutta mallia, mutta laajentaa sitä kokoamalla tiedonhallinnan ja tietämyksen esittämisen, tiedon louhinnan työnkulun sekä ohjelmistoarkkitehtuurin osatekijöiksi yhdistettyyn malliin. Lisäksi malli kattaa aiheita, joita tavallisesti ei oteta huomioon tai joiden ei katsota kuuluvan osaksi tiedonlouhintaprosessia; tällaisia ovat esimerkiksi tiedon louhintaan liittyvät filosofiset kysymykset. Väitöskirjassa käsitellään myös kahta ohjelmistokehystä ja neljää tapaustutkimuksena esiteltävää sovellusta, jotka edustavat teknisiä ratkaisuja eräisiin yksittäisiin tiedonlouhintaprosessin osaongelmiin. Kehykset ja sovellukset toteuttavat ja havainnollistavat useita ehdotetun prosessimallin merkittävimpiä ominaisuuksia.

The abundance of data at our disposal empowers data-driven applications and decision making. The knowledge captured in the data, however, has not been utilized to full potential, as it is only accessible to human interpretation and data are distributed in heterogeneous repositories. Ontologies are a key technology unlocking the knowledge in the data by providing means to model the world around us and infer knowledge implicitly captured in the data. As data are hosted by independent organizations we often need to use several ontologies and discover the relationships between them in order to support data and knowledge transfer. Broadly speaking, while ontologies provide formal representations and thus the basis, ontology alignment supplies integration techniques and thus the means to turn the data kept in distributed, heterogeneous repositories into valuable knowledge. While many automatic approaches for creating alignments have already been developed, user input is still required for obtaining the highest-quality alignments. This thesis focuses on supporting users during the cognitively intensive alignment process and makes several contributions. We have identified front- and back-end system features that foster user involvement during the alignment process and have investigated their support in existing systems by user interface evaluations and literature studies. We have further narrowed down our investigation to features in connection to the, arguably, most cognitively demanding task from the users’ perspective—manual validation—and have also considered the level of user expertise by assessing the impact of user errors on alignments’ quality. As developing and aligning ontologies is an error-prone task, we have focused on the benefits of the integration of ontology alignment and debugging. We have enabled interactive comparative exploration and evaluation of multiple alignments at different levels of detail by developing a dedicated visual environment—Alignment Cubes—which allows for alignments’ evaluation even in the absence of reference alignments. Inspired by the latest technological advances we have investigated and identified three promising directions for the application of large, high-resolution displays in the field: improving the navigation in the ontologies and their alignments, supporting reasoning and collaboration between users.

Robôs sociáveis devem ser capazes de interagir, se comunicar, compreender e se relacionar com os seres humanos de uma forma natural. Existem diversas motivações práticas e científicas para o desenvolvimento de robôs sociáveis como plataforma de pesquisas, educação e entretenimento. Entretanto, embora diversos robôs sociáveis já tenham sido desenvolvidos com sucesso, ainda existe muito trabalho para aprimorar a sua eficácia. A utilização de uma arquitetura robótica pode reduzir fortemente o esforço requerido para o desenvolvimento de um robô sociável. Tal arquitetura robótica deve possuir estruturas e mecanismos para permitir a interação social, o controle do comportamento e o aprendizagem a partir do ambiente. Tal arquitetura deve ainda possuir estruturas e mecanismos para permitir a percepção e a atenção, a fim de permitir que um robô sociável perceba a riqueza do comportamento humano e do meio ambiente, e para aprender a partir de interações sociais. Os processos de aprendizado evidenciados na Análise do Comportamento podem levar ao desenvolvimento de métodos e estruturas promissoras para a construção de robôs sociáveis capazes de aprender por meio da interação com o meio ambiente e de exibir comportamento social apropriado. O proposito deste trabalho é o desenvolvimento de uma arquitetura robótica inspirada na Análise do Comportamento. A arquitetura desenvolvida é capaz de simular o aprendizado do comportamento operante e os métodos e estruturas propostos permitem o controlo e a exibição de comportamentos sociais apropriados e o aprendizado a partir da interação com o meio ambiente. A arquitetura proposta foi avaliada no contexto de um problema real não trivial: o aprendizado da atenção compartilhada. Os resultados obtidos mostram que a arquitetura é capaz de exibir comportamentos apropriados durante uma interação social real e controlada. Ainda, os resultados mostram também que a arquitetura pode aprender a partir de uma interação social. Este trabalho é a base para o desenvolvimento de uma ferramenta para a construção dos robôs sociáveis. Os resultados obtidos abrem muitas oportunidades de trabalhos futuros / Sociable robots should be able to interact, to communicate, to understand and to relate with human beings in a natural way. There are several scientific and practical motivations for developing sociable robots as platform of researches, education and entertainment. However, although several sociable robots have already been developed with success, much work remains to increase their effectiveness. The use of a robotic architecture may strongly reduce the time and effort required to construct a sociable robot. Such robotic architecture must have structures and mechanisms to allow social interaction, behavior control and learning from environment. Such architecture must also have structures and mechanisms to allow perception and attention, to enable a sociable robot to perceive the richness of the human behavior and of the environment, and to learn from social interactions. Learning processes evidenced on Behavior Analysis can led to the development of promising methods and structures for the construction social robots that are able to learn through interaction from the environment and to exhibit appropriate social behavior. The purpose of this work is the development of a robotic architecture inspired from Behavior Analysis. The developed architecture is able to simulate operant behavior learning and the proposed methods and structures allow the control and exhibition of appropriate social behavior and learning from interaction in the environment. The proposed architecture was evaluated in the context of a non trivial real problem: the learning of the shared attention. The obtained results show that the architecture is able to exhibit appropriate behaviors during a real and controlled social interaction. Additionally, the results show also that the architecture can learn from a social interaction. This work is the basis for developing a tool for the construction of social robots. The obtained results open oportunities of many future works

Os Sistemas de Apoio à Decisão (SAD) organizam e processam dados e informações para gerar resultados que apoiem a tomada de decisão em um domínio especifico. Eles integram conhecimento de especialistas de domínio em cada um de seus componentes: modelos, dados, operações matemáticas (que processam os dados) e resultado de análises. Nas metodologias de desenvolvimento tradicionais, esse conhecimento deve ser interpretado e usado por desenvolvedores de software para implementar os SADs. Isso porque especialistas de domínio não conseguem formalizar esse conhecimento em um modelo computável que possa ser integrado aos SADs. O processo de modelagem de conhecimento é realizado, na prática, pelos desenvolvedores, parcializando o conhecimento do domínio e dificultando o desenvolvimento ágil dos SADs (já que os especialistas não modificam o código diretamente). Para solucionar esse problema, propõe-se um método e ferramenta web que usa ontologias, na Web Ontology Language (OWL), para representar o conhecimento de especialistas, e uma Domain Specific Language (DSL), para modelar o comportamento dos SADs. Ontologias, em OWL, são uma representação de conhecimento computável, que permite definir SADs em um formato entendível e accessível a humanos e máquinas. Esse método foi usado para criar o Framework Decisioner para a instanciação de SADs. O Decisioner gera automaticamente SADs a partir de uma ontologia e uma descrição naDSL, incluindo a interface do SAD (usando uma biblioteca de Web Components). Um editor online de ontologias, que usa um formato simplificado, permite que especialistas de domínio possam modificar aspectos da ontologia e imediatamente ver as consequência de suasmudanças no SAD.Uma validação desse método foi realizada, por meio da instanciação do SAD SustenAgro no Framework Decisioner. O SAD SustenAgro avalia a sustentabilidade de sistemas produtivos de cana-de-açúcar na região centro-sul do Brasil. Avaliações, conduzidas por especialistas em sustentabilidade da Embrapa Meio ambiente (parceiros neste projeto), mostraram que especialistas são capazes de alterar a ontologia e DSL usadas, sem a ajuda de programadores, e que o sistema produz análises de sustentabilidade corretas. / Decision Support Systems (DSSs) organize and process data and information to generate results to support decision making in a specific domain. They integrate knowledge from domain experts in each of their components: models, data, mathematical operations (that process the data) and analysis results. In traditional development methodologies, this knowledge must be interpreted and used by software developers to implement DSSs. That is because domain experts cannot formalize this knowledge in a computable model that can be integrated into DSSs. The knowledge modeling process is carried out, in practice, by the developers, biasing domain knowledge and hindering the agile development of DSSs (as domain experts cannot modify code directly). To solve this problem, a method and web tool is proposed that uses ontologies, in the Web Ontology Language (OWL), to represent experts knowledge, and a Domain Specific Language (DSL), to model DSS behavior. Ontologies, in OWL, are a computable knowledge representations, which allow the definition of DSSs in a format understandable and accessible to humans and machines. This method was used to create the Decisioner Framework for the instantiation of DSSs. Decisioner automatically generates DSSs from an ontology and a description in its DSL, including the DSS interface (using a Web Components library). An online ontology editor, using a simplified format, allows that domain experts change the ontology and immediately see the consequences of their changes in the in the DSS. A validation of this method was done through the instantiation of the SustenAgro DSS, using the Decisioner Framework. The SustenAgro DSS evaluates the sustainability of sugarcane production systems in the center-south region of Brazil. Evaluations, done by by sustainability experts from Embrapa Environment (partners in this project), showed that domain experts are capable of changing the ontology and DSL program used, without the help of software developers, and that the system produced correct sustainability analysis.

<div>Deep learning models have become the de-facto solution in various fields like computer vision, natural language processing, robotics, drug discovery, and many others. The skyrocketing performance and success of multi-layer neural networks comes at a significant power and energy cost. Thus, there is a need to rethink the current trajectory and explore different computing frameworks. One such option is spiking neural networks (SNNs) that is inspired from the spike-based processing observed in biological brains. SNNs operating with binary signals (or spikes), can potentially be an energy-efficient alternative to the power-hungry analog neural networks (ANNs) that operate on real-valued analog signals. The binary all-or-nothing spike-based communication in SNNs implemented on event-driven hardware offers a low-power alternative to ANNs. A spike is a Delta function with magnitude 1. With all its appeal for low power, training SNNs efficiently for high accuracy remains an active area of research. The existing ANN training methodologies when applied to SNNs, results in networks that have very high latency. Supervised training of SNNs with spikes is challenging (due to discontinuous gradients) and resource-intensive (time, compute, and memory).Thus, we propose compression methods, training methodologies, learning rules</div><div><br></div><div>First, we propose compression techniques for SNNs based on unsupervised spike timing dependent plasticity (STDP) model. We present a sparse SNN topology where non-critical connections are pruned to reduce the network size and the remaining critical synapses are weight quantized to accommodate for limited conductance levels in emerging in-memory computing hardware . Pruning is based on the power law weight-dependent</div><div>STDP model; synapses between pre- and post-neuron with high spike correlation are retained, whereas synapses with low correlation or uncorrelated spiking activity are pruned. The process of pruning non-critical connections and quantizing the weights of critical synapses is</div><div>performed at regular intervals during training.</div><div><br></div><div>Second, we propose a multimodal SNN that combines two modalities (image and audio). The two unimodal ensembles are connected with cross-modal connections and the entire network is trained with unsupervised learning. The network receives inputs in both modalities for the same class and</div><div>predicts the class label. The excitatory connections in the unimodal ensemble and the cross-modal connections are trained with STDP. The cross-modal connections capture the correlation between neurons of different modalities. The multimodal network learns features of both modalities and improves the classification accuracy compared to unimodal topology, even when one of the modality is distorted by noise. The cross-modal connections are only excitatory and do not inhibit the normal activity of the unimodal ensembles. </div><div><br></div><div>Third, we explore supervised learning methods for SNNs.Many works have shown that an SNN for inference can be formed by copying the weights from a trained ANN and setting the firing threshold for each layer as the maximum input received in that layer. These type of converted SNNs require a large number of time steps to achieve competitive accuracy which diminishes the energy savings. The number of time steps can be reduced by training SNNs with spike-based backpropagation from scratch, but that is computationally expensive and slow. To address these challenges, we present a computationally-efficient training technique for deep SNNs. We propose a hybrid training methodology:</div><div>1) take a converted SNN and use its weights and thresholds as an initialization step for spike-based backpropagation, and 2) perform incremental spike-timing dependent backpropagation (STDB) on this carefully initialized network to obtain an SNN that converges within few epochs and requires fewer time steps for input processing. STDB is performed with a novel surrogate gradient function defined using neuron’s spike time. The weight update is proportional to the difference in spike timing between the current time step and the most recent time step the neuron generated an output spike.</div><div><br></div><div>Fourth, we present techniques to further reduce the inference latency in SNNs. SNNs suffer from high inference latency, resulting from inefficient input encoding, and sub-optimal settings of the neuron parameters (firing threshold, and membrane leak). We propose DIET-SNN, a low-latency deep spiking network that is trained with gradient descent to optimize the membrane leak and the firing threshold along with other network parameters (weights). The membrane leak and threshold for each layer of the SNN are optimized with end-to-end backpropagation to achieve competitive accuracy at reduced latency. The analog pixel values of an image are directly applied to the input layer of DIET-SNN without the need to convert to spike-train. The first convolutional layer is trained to convert inputs into spikes where leaky-integrate-and-fire (LIF) neurons integrate the weighted inputs and generate an output spike when the membrane potential crosses the trained firing threshold. The trained membrane leak controls the flow of input information and attenuates irrelevant inputs to increase the activation sparsity in the convolutional and dense layers of the network. The reduced latency combined with high activation sparsity provides large improvements in computational efficiency.</div><div><br></div><div>Finally, we explore the application of SNNs in sequential learning tasks. We propose LITE-SNN, a lightweight SNN suitable for sequential learning tasks on data from dynamic vision sensors (DVS) and natural language processing (NLP). In general sequential data is processed with complex recurrent neural networks (like long short-term memory (LSTM), and gated recurrent unit (GRU)) with explicit feedback connections and internal states to handle the long-term dependencies. Whereas neuron models in SNNs - integrate-and-fire (IF) or leaky-integrate-and-fire (LIF) - have implicit feedback in their internal state (membrane potential) by design and can be leveraged for sequential tasks. The membrane potential in the IF/LIF neuron integrates the incoming current and outputs an event (or spike) when the potential crosses a threshold value. Since SNNs compute with highly sparse spike-based spatio-temporal data, the energy/inference is lower than LSTMs/GRUs. SNNs also have fewer parameters than LSTM/GRU resulting in smaller models and faster inference. We observe the problem of vanishing gradients in vanilla SNNs for longer sequences and implement a convolutional SNN with attention layers to perform sequence-to-sequence learning tasks. The inherent recurrence in SNNs, in addition to the fully parallelized convolutional operations, provides an additional mechanism to model sequential dependencies and leads to better accuracy than convolutional neural networks with ReLU activations.</div>

<div>Autonomous transportation is a research field that has gained huge interest in recent years, with autonomous electric or hydrogen cars coming ever closer to seeing everyday use. Not just cars are subject to autonomous research though, the field of aviation is also being explored for fully autonomous flight. One very important aspect for making autonomous flight a reality is attitude control, the control of roll, pitch, and sometimes yaw. Traditional approaches for automated attitude control use PID (proportional-integral-derivative) controllers, which use hand-tuned parameters to fulfill the task. In this work, however, the use of Reinforcement Learning algorithms for attitude control will be explored. With the surge of more and more powerful artificial neural networks, which have proven to be universally usable function approximators, Deep Reinforcement Learning also becomes an intriguing option. </div><div>A software toolkit will be developed and used to allow for the use of multiple flight simulators to train agents with Reinforcement Learning as well as Deep Reinforcement Learning. Experiments will be run using different hyperparamters, algorithms, state representations, and reward functions to explore possible options for autonomous attitude control using Reinforcement Learning.</div>

E-scooters are ubiquitous and their number keeps escalating, increasing their interactions with other vehicles on the road. E-scooter riders have an atypical behavior that varies enormously from other vulnerable road users, creating new challenges for vehicle active safety systems and automated driving functionalities. The detection of e-scooter riders by other vehicles is the first step in taking care of the risks. This research presents a novel vision-based system to differentiate between e-scooter riders and regular pedestrians and a benchmark dataset for e-scooter riders in natural environments. An efficient system pipeline built using two existing state-of-the-art convolutional neural networks (CNN), You Only Look Once (YOLOv3) and MobileNetV2, performs detection of these vulnerable e-scooter riders.<br>