JKJoon Kwon

Joon Kwon

*This is an algorithmically generated author record based on publicly available information.

Current Organizations / Institutes

The organizations/institutes from the scholar's latest article

Previous Organizations / Institutes

The organizations / institutes from the scholar's other articles in the past ten years


Joon Kwon has a total of 24 co-authors. Here are the top co-authors by number of shared publications:


Latest Publications

  • Mirror descent strategies for regret minimization and approachability

    by Joon Kwon

    Stratégies de descente miroir pour la minimisation du regret et l'approchabilité: On présente dans le Chapitre I le problème d'online linear optimization, et on étudie les stratégies de descente miroir. Le Chapitre II se concentre sur le cas où le joueur dispose d'un ensemble fini d'actions. Le Chapitre III établit que les stratégies FTPL appartiennent à la famille de descente miroir. On construit au Chapitre IV des stratégies de descente miroir pour l'approchabilité de Blackwell. Celles-ci sont ensuite appliquées à construction de stratégies optimales pour le problème online combinatorial optimization et la minimisation du regret interne/swap. Le Chapitre V porte sur la minimisation du regret avec l'hypothèse supplémentaire que les vecteurs de paiement possèdent au plus s composantes non-nulles. On met en évidence une différence fondamentale entre les gains et les pertes en établissant des bornes optimales sur le regret d'ordre différents dans chacun de ces deux cas. Le Chapitre VI porte sur l'approchabilité de Blackwell avec observations partielles. On établit que les vitesses de convergence optimales sont O(T−1/2) pour des signaux dont les lois ne dépendent pas de l'action du joueur, et O(T−1/3) dans le cas général. Le Chapitre VII définit les stratégies de descente miroir en temps continu. On établit pour ces derniers une propriété de non-regret. On effectue ensuite une comparaison entre le temps continu et le temps discret. Enfin, le Chapitre VIII établit une borne universelle sur les variations des fonctions convexes bornées. On obtient en corollaire que toute fonction convexe bornée est lipschitzienne par rapport à la métrique de Hilbert.

  • A pipeline with pre-processing options to detect behaviour from accelerometer data using Machine Learning tested on dairy goats

    by Sarah Mauny, Joon Kwon, Nicolas C. Friggens, Christine Duvaux-Ponter, Masoomeh Taghipoor

    Animal behaviour is a significant component in the evaluation of animal welfare. Conducting continuous observations of animal behaviour is a time-consuming task and may not be feasible over extended periods for all animals. Thus, new technologies like sensors and cameras can be used to assess individual behaviour continuously. Combined with Artificial Intelligence (AI), accelerometers are promising to continuously and individually detect animal behaviour from the acceleration signals and characteristics of the behaviour. Such devices are commercialised for cattle but they have not been widely developed for small ruminants. Being able to automatically monitor behaviour at an individual scale represents a crucial step towards an objective assessment of animal welfare. This paper aims to present the use of a pipeline called ACT4Behav (Accelerometer-based Classification Tool for identifying Behaviours) involving a supervised classification algorithm for automatically characterising specific animal behaviours using accelerometer data, and to explore the best pre-processing steps for each behaviour. This algorithm is designed to be general-purpose and applicable with different species, behaviours and accelerometers. This paper presents the use of this pipeline with eight indoor-housed goats equipped with ear-mounted accelerometers. Rumination, head in the feeder, standing and lying behaviours were continuously sampled from camera recordings for 11 consecutive hours for each goat using The Observer software. The developed pipeline was used to identify optimal descriptive features and data preparation steps for each prediction model, one for each behaviour. A sensitivity analysis was conducted to assess the impact of the processing techniques and parameter value on the resulting AUC (Area Under the Curve) score, used as the performance score of the models. This analysis allowed the identification of the adequate filtering techniques, time-window segmentations, application of various transformations to raw data, and feature selections for each behaviour. Tuning the data pre-processing for each behaviour enhanced the ability to predict rumination (AUC score=0.800), head in the feeder (AUC score=0.819), lying (AUC score=0.829) and standing (AUC score=0.823) behaviours. When the application of the models on goats that did not participate in the training was tested by training the models on six goats and testing it on the two other goats, the AUC score for the four behaviours decreased (0.644, 0.733, 0.741 and 0.749 respectively for rumination, head in the feeder, lying and standing).


  • Open Access

    Importance sampling-based gradient method for dimension reduction in Poisson log-normal model

    by Bastien Batardière, Julien Chiquet, Joon Kwon, Julien Stoehr

    High-dimensional count data poses significant challenges for statistical analysis, necessitating effective methods that also preserve explainability. We focus on a low rank constrained variant of the Poisson log-normal model, which relates the observed data to a latent low-dimensional multivariate Gaussian variable via a Poisson distribution. Variational inference methods have become a golden standard solution to infer such a model. While computationally efficient, they usually lack theoretical statistical properties with respect to the model. To address this issue we propose a projected stochastic gradient scheme that directly maximizes the log-likelihood. We prove the convergence of the proposed method when using importance sampling for estimating the gradient. Specifically, we achieve a convergence rate of O(T−1∕2+N−1), where T denotes the number of iterations and N represents the number of Monte Carlo samples. The latter follows from a novel descent lemma for non convex L-smooth objective functions, and random biased gradient estimate. We also demonstrate numerically the efficiency of our solution compared to its variational competitor. Our method not only scales with respect to the number of observed samples but also provides access to the desirable properties of the maximum likelihood estimator.


  • Open Access

    Finite-sum optimization: Adaptivity to smoothness and loopless variance reduction

    by Bastien Batardière, Joon Kwon

    For finite-sum optimization, variance-reduced gradient methods (VR) compute at each iteration the gradient of a single function (or of a mini-batch), and yet achieve faster convergence than SGD thanks to a carefully crafted lower-variance stochastic gradient estimator that reuses past gradients. Another important line of research of the past decade in continuous optimization is the adaptive algorithms such as AdaGrad, that dynamically adjust the (possibly coordinate-wise) learning rate to past gradients and thereby adapt to the geometry of the objective function. Variants such as RMSprop and Adam demonstrate outstanding practical performance that have contributed to the success of deep learning. In this work, we present AdaLVR, which combines the AdaGrad algorithm with loopless variance-reduced gradient estimators such as SAGA or L-SVRG that benefits from a straightforward construction and a streamlined analysis. We assess that AdaLVR inherits both good convergence properties from VR methods and the adaptive nature of AdaGrad: in the case of $L$-smooth convex functions we establish a gradient complexity of $O(n+(L+\sqrt{nL})/\varepsilon)$ without prior knowledge of $L$. Numerical experiments demonstrate the superiority of AdaLVR over state-of-the-art methods. Moreover, we empirically show that the RMSprop and Adam algorithm combined with variance-reduced gradients estimators achieve even faster convergence.

  • Unifying mirror descent and dual averaging

    by Anatoli Juditsky, Joon Kwon, Éric Moulines


  • Open Access

    Global Plant Extinction Risk Assessment Inform Novel Biodiversity Hotspots

    by Thomas Haevermans, Jessica Tressou, Joon Kwon, Roseli Pellens, Anne Dubéarnès, Simon Veron, Liliane Bel, Stéphane Dervaux, Juliette Dibie-Barthelemy, Myriam Gaudeul, et al.

    Curbing biodiversity loss and its impact on ecosystem services, resilience and Nature’s Contributions to People is one of the main challenges of our generation (IPBES, 2019b, 2019a; Secretariat of the United Nations Convention on Biological Diversity, 2020). A global baseline assessment of the threat status of all of biodiversity is crucial to monitor the progress of conservation policies worldwide (Mace & al., 2000; Secretariat of the United Nations Convention on Biological Diversity, 2021) and target priority areas for conservation (Walker & al., 2021). However, the magnitude of the task seems insurmountable, as even listing the organisms already known to science is a challenge (Nic Lughadha & al., 2016; Borsch & al., 2020; Govaerts & al., 2021). A new approach is needed to overcome this stumbling block and scale-up the assessment of extinction risk. Here we show that analyses of natural history mega-datasets using artificial intelligence allows us to predict a baseline conservation status for all vascular plants and identify target areas for conservation corresponding to hotspots optimally capturing different aspects of biodiversity. We illustrate the strong potential of AI-based methods to reliably predict extinction risk on a global scale. Our approach not only retrieved recognized biodiversity hotspots but identified new areas that may guide future global conservation action (Myers & al., 2000; Brooks & al., 2006). To further work in this area and guide the targets of the post-2020 biodiversity framework (Díaz & al., 2020a; Secretariat of the United Nations Convention on Biological Diversity, 2020; Mair & al., 2021), it will be necessary to accelerate the acquisition of fundamental data and allow inclusion of social and economic factors (Possingham & Wilson, 2005).


  • Open Access

    Refined approachability algorithms and application to regret minimization with global costs

    by Joon Kwon

    Blackwell's approachability is a framework where two players, the Decision Maker and the Environment, play a repeated game with vector-valued payoffs. The goal of the Decision Maker is to make the average payoff converge to a given set called the target. When this is indeed possible, simple algorithms which guarantee the convergence are known. This abstract tool was successfully used for the construction of optimal strategies in various repeated games, but also found several applications in online learning. By extending an approach proposed by (Abernethy et al., 2011), we construct and analyze a class of Follow the Regularized Leader algorithms (FTRL) for Blackwell's approachability which are able to minimize not only the Euclidean distance to the target set (as it is often the case in the context of Blackwell's approachability) but a wide range of distance-like quantities. This flexibility enables us to apply these algorithms to closely minimize the quantity of interest in various online learning problems. In particular, for regret minimization with $\ell_p$ global costs, we obtain the first bounds with explicit dependence in $p$ and the dimension $d$.