RESEARCH & INNOVATION

SEMINARS
PUBLICATIONS
THESES

VIDEO RECORDS

LIVIA Montreal

LIVIA Montreal
by Reda Bensaid, PhD candidate, Polytechnique Montréal & IMT Atlantique Abstract: Over the last decade, the de facto standard for training high-performing representation learning models has heavily relied on large-scale static datasets crawled from the Internet. However, recent advances in visual content creation challenge this status quo by pushing researchers to leverage high-performing image-generative models as world models that work in tandem with representation learning models by providing them with data. In this talk, I will address two research questions: (1) Are state-of-the-art image generative models optimized to work as world models? and (2) What is the most effective way to guide the generative model to produce samples useful for the downstream task? Bio: My name is Reda Bensaid. I am currently in my 2nd year of pursuing a PhD at Polytechnique Montréal & IMT Atlantique, where I specialize in compressing and adapting vision foundation models for few-shot semantic segmentation. Originally from Morocco, I studied math and physics before earning my engineering degree from IMT Atlantique in France. Prior to starting my PhD, I worked as a data scientist. Find out more information related to our research at the LIVIA website: https://liviamtl.ca/
by Quentin BERTRAND, post-doctoral researcher, MILA Abstract: Over the last decade, the de facto standard for training high performing representation learning models has heavily relied on large scale static datasets crawled from the Internet. However, recent advances in visual content creation are challenging this status quo by pushing researchers to leverage high performing image generative models as world models that work in tandem with representation learning models by providing them with data. In this talk, I will address two research questions: (1) Are state-of-the-art image generative models optimized to work as world models?; and (2) What is the most effective way to guide the generative model to produce samples that are useful for the downstream task? Bio: Adriana Romero-Soriano is currently a research scientist at Meta AI (FAIR), an adjunct professor at McGill University, a core industry member of Mila, and a Canada CIFAR Chair. The playground of her research has been defined by problems which require inferring full observations from limited sensory data, building models of the world with the goal to improve impactful downstream applications responsibly. Her most recent research focuses on improving the quality, consistency, and representation diversity of visual content creation systems. Adriana received her Ph.D. from University of Barcelona, where she worked with Dr. Carlo Gatta, and spent two years as post-doctoral researcher at Mila working with Prof. Yoshua Bengio. Find out more information related to our research at the LIVIA website: https://liviamtl.ca/
by Masih Aminbeidokhti, PhD candidate, LIVIA-ETS Abstract: Data augmentation is one of the most effective techniques for regularizing deep learning models and improving recognition performance in various tasks and domains. However, this holds for standard in-domain settings where the training and test data follow the same distribution. The best recipe for data augmentation is unclear for the out-of-domain case, where the test data follow a different and unknown distribution. In this paper, we show that for out-of-domain and domain generalization settings, data augmentation can provide a conspicuous and robust improvement in performance. To do that, we propose a simple training procedure: (i) use uniform sampling on standard data augmentation transformations, (ii) increase the strength transformations to account for the higher data variance expected when working out-of-domain, and (iii) devise a new reward function to reject extreme transformations that can harm the training. With this procedure, our data augmentation scheme achieves a level of accuracy comparable to or better than state-of-the-art methods on benchmark domain generalization datasets. Paper available at: https://arxiv.org/pdf/2310.06670.pdf Find out more information related to our research at the LIVIA website: https://liviamtl.ca/
by Heitor Rapela Medeiros, PhD candidate, LIVIA-ETS Abstract: Image translation is a powerful way to adapt a visual recognition model to a new domain. However, common image translation approaches only focus on generating data from the same distribution as the target domain. Given a cross-modal application, such as pedestrian detection from aerial images,  with a considerable shift in data distribution between infrared (IR) to visible (RGB) images, a translation focused on generation might lead to poor performance as the loss focuses on irrelevant details for the task. In this paper, we propose HalluciDet, an IR-RGB image translation model for object detection. Instead of focusing on reconstructing the original image on the IR modality, it seeks to reduce the detection loss of an RGB detector, avoiding the need to access RGB data. This model produces a new image representation that enhances objects of interest in the scene and dramatically improves detection performance. We empirically compare our approach against state-of-the-art methods for image translation and for fine-tuning on IR, and show that our HalluciDet improves detection accuracy in most cases by exploiting the privileged information encoded in a pre-trained RGB detector. Paper available at: https://arxiv.org/pdf/2310.04662.pdf Find out more information related to our research at the LIVIA website: https://liviamtl.ca/
by Jean Paul Barddal, Professor, Pontifícia Universidade Católica do Paraná Abstract: Processing and analyzing potentially unbounded data sequences, the so-called data streams, is key in multiple applications. However, assuming such streams as stationary is bound to fail. In this talk, we will discuss what data stream mining are and what are the fundamental challenges they bring to the table, including learning over time, computational restrictions, and concept drift. Finally, I will bring forward some of the ongoing research we have in our research group including green data stream mining, concept drift adaptation in image classification, and textual data streams. Bio: Jean Paul Barddal received his Ph.D. in Informatics from the Graduate Program in Informatics (PPGIA), in 2018, where he is currently a professor. He is also an executive member of Center for Artificial Intelligence Solutions (CISIA) and the Artificial Intelligence Institute (AI2). His research topics include machine learning from data streams, including classification, regression, clustering, and recommender systems. He has published over 60 papers including developments to algorithms in data stream mining or their application to real-world problems. Find out more information related to our research at the LIVIA website: https://liviamtl.ca/
by Quentin BERTRAND, post-doctoral researcher, MILA Abstract: For several years now, deep learning techniques have been successfully applied to medical imaging in general, and echocardiography in particular. Far from being the promised panacea, these methods have allowed major advances in many specific areas such as image acquisition, echocardiogram analysis, segmentation and tracking of anatomical structures or even user guidance. Moreover, the combination of these different techniques allows the deployment of complete and fully automated processing chains, making it possible to increase the reliability of clinical measurements and facilitate the use of ultrasound scanners outside of hospitals. In my talk, I will describe our last contributions on the segmentation of echocardiographic sequences for the automatic estimation of clinical indices. Our ultimate goal is to develop a methodology to definitively solve this 30-year-old problem. This currently includes designing appropriate databases, exploiting the VAE formalism and modeling uncertainty through the CLIP framework. Bio: Olivier Bernard received his Electrical Engineering degree and Ph.D. from the University of Lyon (INSA), France, in 2003 and 2006, respectively. He was a Postdoctoral Fellow with the Biomedical Imaging Group at the Federal Polytechnic Institute of Lausanne, EPFL, Switzerland in 2007. Currently, he is a Professor with the University of Lyon (INSA) and the CREATIS laboratory in France. He is also the head of the Myriad research team, which specializes in medical image analysis, simulation, and modeling. His current research interests focus on image analysis through deep learning techniques, with applications in cardiovascular imaging, blood flow imaging, and population representation. Prof. Bernard was also an Associate Editor of the IEEE Transactions on Image Processing. Find out more information related to our research at the LIVIA website: https://liviamtl.ca/
by Clemens Pollak, Research Associate, Deep-MI Lab, German Center for Neurodegenerative Diseases Abstract: In the age of deep learning the accuracy and speed of brain segmentation is ever increasing. In the medical imaging domain, however, high average accuracy on pre-defined datasets is not good enough. Large population studies require performant tools that avoid or reduce potential biases in downstream analysis and can be applied to data from all participants, indifferent of large anatomical or pathological variance. We will approach this challenge, first by quantifying motion artifacts, caused by head motion during the MR acquisition, to reduce detrimental biases in analysis of brain changes; and second, by improving the performance of deep learning-based brain segmentation in the presence of large lesions, by generating healthy looking priors. Bio: Pierre LE JEUNE is a PhD student at L2TI laboratory, University Sorbonne Paris Nord while working at COSE company. He received the M.Sc. degree in Mathematical Modelling and Computation from Danish Technical University (Copenhagen) and the M.Sc. in engineering from Centrale Nantes. His current research interests include Few-Shot Learning, Computer Vision and Deep Learning Paper available at: https://arxiv.org/abs/2306.11982 Find out more information related to our research at the LIVIA website: https://liviamtl.ca/
by Mehraveh Javan Roshtkhari, student at the LIVIA Abstract: Down sampling layers, including pooling and strided convolutions, are crucial components of the convolutional neural network architecture that determine both the granularity/scale of image feature analysis as well as the receptive field size of a given layer. To fully understand this problem, we analyse the performance of models independently trained with each pooling configurations and show that the position of the down sampling layers can highly influence the performance of a network and predefined down sampling configurations are not optimal. Network Architecture Search (NAS) might be used to optimize down sampling configurations as an hyperparameter. However, we find that common one-shot NAS based on a single SuperNet does not work for this problem. Finally, we propose a balanced mixture of SuperNets that automatically associates pooling configurations to different weight models and helps to reduce the weight-sharing and inter-influence of pooling configurations on the SuperNet parameters.. Paper available at: https://arxiv.org/abs/2306.11982 Find out more information related to our research at the LIVIA website: https://liviamtl.ca/
by Ségolène Martin, Ph.D. candidate, CentraleSupélec, Université Paris Saclay, INRIA, France Abstract: Standard few-shot benchmarks are often built upon simplifying assumptions on the query sets, which may not always hold in practice. In particular, for each task at testing time, the classes effectively present in the unlabeled query set are known a priori, and correspond exactly to the set of classes represented in the labeled support set. We relax these assumptions and extend current benchmarks, so that the query-set classes of a given task are unknown, but just belong to a much larger set of possible classes. Our setting could be viewed as an instance of the challenging yet practical problem of extremely imbalanced K-way classification, K being much larger than the values typically used in standard benchmarks, and with potentially irrelevant supervision from the support set. Expectedly, our setting incurs drops in the performances of state-of-the-art methods. Motivated by these observations, we introduce a PrimAl Dual Minimum Description LEngth (PADDLE) formulation, which balances data-fitting accuracy and model complexity for a given few-shot task, under supervision constraints from the support set. Our constrained MDL-like objective promotes competition among a large set of possible classes, preserving only effective classes that befit better the data of a few-shot task. It is hyperparameter free, and could be applied on top of any base-class training. Furthermore, we derive a fast block coordinate descent algorithm for optimizing our objective, with convergence guarantee, and a linear computational complexity at each iteration. Comprehensive experiments over the standard few-shot datasets and the more realistic and challenging i-Nat dataset show highly competitive performances of our method, more so when the numbers of possible classes in the tasks increase. Bio: Ségolène Martin received her M.Sc. degree in mathematics from the Ecole Normale Supérieure Paris-Saclay, France, in 2020. She also obtained the qualification for teachers (Agrégation) in 2019. She is currently a Ph.D. student in CVN lab at CentraleSupéelec, University Paris Saclay, France. Her research focuses on optimization methods for image processing and machine learning. Paper available at: https://openreview.net/pdf?id=j9JL96S8Vl Code available at: https://github.com/SegoleneMartin/PADDLE Find out more information related to our research at the LIVIA website: https://liviamtl.ca/
Load More... Subscribe