• Deep Learning Meetup #17 at Samsung Paris

    Le Centorial

    Dear Deeplearners, The Deep Learning meetup is coming back on 18/07, again at Samsung Paris office. You must be registered on the EventBrite event to be accepted for security reason (https://www.eventbrite.co.uk/e/deep-learning-meetup-17-at-samsung-paris-tickets-64993557480). Bring your ID as it will be checked beforehand. We'll have 3 great speakers: Damien MENIGAUX (Veesion): Veesion is relies on video deep learning to detect theft related gestures in real time in retail stores. Since its inception in 2014, video deep learning has made some important strives. The appearance of large scale video datasets was accompanied by important architectural findings. If traditionnal image deep learning is considered by some as a "solved problem", the temporal dimension in videos still isn't well understood. Let's take a look back at the most ingenious and successful innovations in video deep learning networks. Eloi Zablocki (LIP6 Sorbonne Université): As textual resources are abundant and contain high-level knowledge, linguistic representations can be used to augment capacities of computer vision recognition systems, typically when visual supervision is scarce. We thus focus on the zero-shot learning recognition task which consists in recognizing objects that have never been seen, thanks to linguistic knowledge acquired about the objects beforehand. We present a model for zero-shot recognition that leverages (1) the region of interest, (2) the semantic representations of object labels, and (3) the visual context of an object Thibaut Barroyer (Cardiologs): Cardiologs is a medical technology company committed to transforming cardiac diagnostics by utilising medical-grade deep learning techniques. Developing deep learning models to solve medical problems is challenging: FDA clearance, data annotation, clinical expertise in the team… We will give an overview of the two main algorithms that our solution is built upon: networks for classification (diagnosis) and segmentation (waves identification) of electrocardiograms (ECG). We will then present an application to the detection of the Wolff-Parkinson-White syndrome, responsible of sudden cardiac arrest, and present results of a comparative study we ran with physicians. You can check out previous meetup slides there: https://drive.google.com/open?id=1nvXwUbbhDZQf3LYifZQiGW3eFePxgJb_ Or watch them there: https://www.youtube.com/channel/UCF65w-sGTJfDI3WarFwTpwg

    12
  • Deep Learning Meetup #16 at Samsung Paris

    Le Centorial

    Dear Deeplearners, The Deep Learning meetup is coming back on 30/01. For security reasons you will need to subscribe to the eventbrite event, and bring your ID as it will be checked beforehand: https://www.eventbrite.co.uk/e/deep-learning-paris-meetup-16-at-samsung-le-centorial-tickets-55025027338 Only subscriptions on Eventbrite will be taken into account. We'll have 3 great speakers: Thomas Wolf, Chief Science Officer R&D at Hugging Face Transfer Learning for Natural Language Generation – The Case of Open-Domain Dialog Free-form dialogue systems ("chatbots") are agents that are designed to interact with humans in open conversations. Developing these systems tackles the general research question of how a model can generate a coherent text output from a textual input, in particular over a wide range of topics and in a stochastic environment. These dialog agents are thus test-beds for many interactive AI systems but as of today, building such intelligent conversational agents remains an unsolved problem. In this talk, I will present a comparison and the technical details behind the two winning approaches of the Conversational Intelligence Challenge 2 held at NeurIPS 2018 last month. The first (our) approach won the automatic evaluation track while the second approach won the human evaluation track. These two approach's bears interesting similarities, being both based on a transfer learning scheme and more precisely on the very same pre-trained model, but also showcase interesting and complementary differences in the implementation of the adaptation phase with differing fine-tuning datasets, multi-task objectives and architectural adaptations. Mathieu Poumeyrol, Senior R&D at Snips Tract: running deep models on the edge Tract is Snips' Open Source Neural Network inference library. It can run the "Hey Snips" wakeword model or Google Inception on a 5$ chip. Alexandre Ramé, Lead R&D at Heuritech OMNIA Faster RCNN (https://arxiv.org/abs/1812.02611) Object detectors tend to perform poorly in new or open domains, and require exhaustive yet costly annotations from fully labeled datasets. We aim at benefiting from several datasets with different categories but without additional labelling, not only to increase the number of categories detected, but also to take advantage from transfer learning and to enhance domain independence. More details about the talks later. Looking forward to seeing you all there ! Charles Ollion

    18
  • Deep Learning Meetup #15 at Palais des Congrès

    Le Palais des Congres de Paris

    Dear Deeplearners, On November 06 will take place our 15th Deep Learning Meetup, organised by Heuritech in collaboration with Microsoft. It will take place at Palais des Congrès, after Microsoft Experience day 1. We'll have 3 great speakers, starting with Andrew Fitzgibbon (Microsoft), Léonard Blier (FAIR / TAU INRIA) and Pierre Stefani (Photobox), followed by informal discussion, food & drinks. ** VERY IMPORTANT ** for security reasons at Palais des Congrès the admission is strictly limited to people subscribed to the event Microsoft Experiences. It is absolutely free and you may register there just for the meetup: https://experiences18.microsoft.fr/ . Moreover, the gates will close at 18:30 strictly so please come before. If you realise you won't be able to join, please RSVP as no on Meetup. == Andrew Fitzgibbon (Microsoft) I have been lucky enough to have been involved in the development of real-world computer vision systems for over twenty years. In 1999, prize-winning research from Oxford University was spun out to become the Emmy-award-winning camera tracker “boujou”, which has been used to insert computer graphics into live-action footage in pretty much every movie made since its release, from the “Harry Potter” series to “Bridget Jones’s Diary”. In 2007, I was part of the team that delivered human body tracking in Kinect for Xbox360, and in 2015 I moved from Microsoft Research to the Windows division to work on Microsoft’s HoloLens, an AR headset brimming with cutting-edge computer vision technology. In all of these projects, the academic state of the art has had to be leapfrogged in accuracy and efficiency, sometimes by several orders of magnitude. Sometimes that’s just raw engineering, sometimes it means completely new ways of looking at the research. If I had to nominate one key to success, it’s a focus on, well, everything: from low-level coding to algorithms to user interface design, and on always being willing to change one’s mind. Andrew is a scientist with HoloLens at Microsoft, Cambridge. He is best known for his work on 3D vision, computer vision, graphics, machine learning, and a little neuroscience. He has published numerous highly-cited papers, and received many awards for his work, including ten “best paper” prizes at various venues, the Silver medal of the Royal Academy of Engineering, and the BCS Roger Needham award. == Léonard Blier, (FAIR Paris / TAU INRIA) - Do Deep Learning Models have too many parameters?" "The best model is the simplest model which can explain the data", says Occam's razor principle, which describes how we are doing inductive reasoning and generalization, both in empirical science and in our everyday life. This general principle can be formalized with tools from information theory and compression: the best model is the model which can compress (losslessly) the data the most when taking into account the cost of encoding the model itself. When it comes to Deep Learning, there seems to be a paradox: in practice, the best models are often huge, with a lot of parameters, so extremely expensive to encode. We will introduce the information theory viewpoint in machine learning and deep learning, and show that despite their huge number of parameters, deep learning models are not "complex". == Pierre Stefani (Photobox) - Landmark recognition in photos Where did I took this photo ? What was the name of that castle again ? Recognizing landmarks and touristic sites in pictures is not a straightforward task: we’ll show the main challenges that we faced to tackle this problem, available resources, including the recent Google Landmarks Dataset and their Deep Attentive Local Features. Since off-the shelf solutions trained on public data do not perform as well for Photobox usecases, we will present our adapted solution that combines fine-tuned CNNs with our own datasets, an attention model and locally aggregated vectors (VLAD). Looking forward to see you all Charles

    2
  • Deep Learning Meetup #14 at Station F

    Station F

    English version below Chers Deeplearners, Le meetup Deep Learning revient le 26/09. L'événement sera spécial cette fois, car nous aurons l'occasion de nous retrouver à Station F à 19h00, et que le Meetup est organisé en partenariat entre Heuritech et Microsoft. Il sera suivi de discussions autour de pizzas et boissons vers 21h. Pour s'inscrire il faut impérativement s'inscrire au lien suivant en indiquant nom et prénom qui seront vérifiés à Station F (Être inscrit sur Meetup ne vous garantit en aucun cas l'accès): http://bit.ly/EventbriteHeuritech Nous aurons l'honneur d'avoir Gül Varol et Julien Perez, chercheuse et chercheur en Deep Learning. Exceptionnellement, nous aurons également Lê Nguyên Hoang, vidéaste Youtube responsable de la chaîne Science4all. Les abstracts de leurs présentations en fin de message. Au plaisir de vous y voir, Charles -- Dear Deeplearners, The Deep Learning meetup is coming back on 26/09. The event will have a special flavor, as we will meet at Station F at 19:00, and because the meetup is co-organized between Heuritech and Microsoft. Discussions, Pizzas & Drinks will follow at around 21:00. To subscribe, you absolutely need to register using the following link with your name and surname which will be checked at Station F (Being registered here on Meetup does not guarantee any access): http://bit.ly/EventbriteHeuritech We will be honored to have Gül Varol and Julien Perez, researchers in Deep Learning. Exceptionnally, we will also have a talk from Lê Nguyên Hoang, the French Youtuber who created the Science4all channel (Talk in French). You may find the abstracts of their presentations below. Best, Charles -- Gul Varol (INRIA/ENS) BodyNet: Volumetric Inference of 3D Human Body Shapes - English Human shape estimation is an important task for video editing, animation and fashion industry. Predicting 3D human body shape from natural images, however, is highly challenging due to factors such as variation in human bodies, clothing and viewpoint. We propose BodyNet, an end-to-end trainable neural network for direct inference of volumetric body shape from a single image. First, I will present our recently released SURREAL dataset that consists of synthetic images of people whose 3D annotations are automatically collected. Then, I will show the advantages of the BodyNet components: (i) a volumetric 3D loss, (ii) a multi-view re-projection loss, and (iii) intermediate supervision of 2D pose, 2D body part segmentation, and 3D pose. Our results and the dataset open up new possibilities for advancing person analysis using cheap and large-scale synthetic data. More at: https://www.di.ens.fr/willow/research/bodynet/ -- Julien Perez (Naver Labs) Machine Reading - French or English The field of Machine Reading has recently emerged as a possible continuation of the tasks of Natural Language Processing. Given a large set of passages of text associated with questions and answers, our goal consists in learning a question answering system solely from these examples. As the research groups dedicated to Machine Reading around the globe have started to produce encouraging results, this task challenges our current understanding of deep learning and machine comprehension. In this talk, we will go through some of the current models and learning protocols associated with this task. After describing the current datasets of the domain, we will introduce ReviewQA, a dataset of relational reasoning for review understanding. Finally, we will discuss other possible applications of such an approach like dialog understanding and fact news detection. -- Lê Nguyên Hoang (Science4all Youtube Channel) Faut-il craindre une superintelligence artificielle ? - French On va étudier les prédictions des experts et d’autres arguments autour de la possibilité d’une superintelligence artificielle, de l’éventuelle date de son émergence, et de ses conséquences.

    8
  • Deep Learning Meetup #13

    Heuritech

    Dear DeepLearners, Our next Meetup is scheduled on Wednesday March 7 at 7:00 pm. This meetup is a little special, since it is a bit more focused into research than usual... To those afraid of math, beware! --------------------------- Olivier Grisel (ML Expert - Software Engineer at INRIA) - Generalization in Deep networks Abstract: This talk will give an overview to some recent theoretical results and experiments on why deep learning models work so well (when they work). In particular we will discuss expressive power, optimization and generalization and their interaction. We will illustrate some of the main insights with empirical experiments. ----------------------------- Arthur Mensch (PhD Candidate at INRIA Parietal) - Differentiable Dynamic Programming for Structured Prediction and Attention Abstract: Dynamic programming (DP) solves a variety of structured combinatorial problems by iteratively breaking them down into smaller subproblems. In spite of their versatility, DP algorithms are usually non-differentiable, which hampers their use as a layer in neural networks trained by backpropagation. To address this issue, we propose to smooth the max operator in the dynamic programming recursion, using a strongly convex regularizer. This allows to relax both the optimal value and solution of the original combinatorial problem, and turns a broad class of DP algorithms into differentiable operators. Theoretically, we provide a new probabilistic perspective on backpropagating through these DP operators, and relate them to inference in graphical models. We derive two particular instantiations of our framework, a smoothed Viterbi algorithm for sequence prediction and a smoothed DTW algorithm for time-series alignment. We showcase these instantiations on two structured prediction tasks and on structured and sparse attention for neural machine translation. ----------------------------- Diogo Luvizon (PhD Candidate at ETIS - Université de Cergy) - 2D/3D Pose Estimation and Action Recognition using Multitask Deep Learning Abstract: Action recognition and human pose estimation are closely related tasks since both problems depend on the human body understanding and additionally, action recognition benefits from precise estimated poses. Despite that, both problems are generally handled as distinct tasks in the literature. In this work, we propose a multitask framework for jointly 2D and 3D pose estimation from still images and human action recognition from video sequences. We show that a single architecture can be used to solve the two problems in an efficient way and still achieves state-of-the-art results. Additionally, we demonstrate that optimization from end-to-end leads to significantly higher accuracy than separated learning. The proposed architecture can be trained with data from different categories simultaneously in a seamlessly way. The reported results on four datasets (MPII, Human3.6M, Penn Action and NTU) demonstrate the effectiveness of our method on the targeted tasks. --------------------------- As usual, pizzas and drinks will be provided after the talks. Heuritech Meetup Team.

    10
  • Deep Learning Meetup #12

    Heuritech

    Dear DeepLearners, Our next Meetup is scheduled on Wednesday December 20 at 7:00 pm Micael Carvalho (PhD candidate at LIP6) - Metric learning for crossmodal alignment Abstract: Modeling different data types into the same representation space is one of the leading approaches for solving problems involving visual and textual data. Recently, large-scale datasets allowed them to be tackled from different angles, including the use of partial annotations. In the light of these advances, we introduce modern metric learning strategies for feature-space alignment, with particular interest in multi-task and cross-modality. --------------------------- Timothée Lacroix (PhD candidate at Facebook AI Research) - Knowledge Base Completion : state of the art and challenges Abstract: We'll start with an introduction to the problem of knowledge base completion and a few methods that have been proposed. Then we'll mention generalization guarantees and regularization for these methods, linking to the work done in collaborative filtering. --------------------------- Renaud Allioux (CTO at EarthCube) - Satellite Imagery and weakly supervised learning Abstract: After introducing specificites of deep learning application using satellite images, we will discuss the problem of image annotation and the application of weakly supervised method to detect small objects --------------------------- As usual, pizzas and drinks will be provided after the talks. Heuritech Meetup Team.

    12
  • Deep Learning Meetup #11

    Heuritech

    Dear DeepLearners, Long time no see ! We hope you had wonderful vacations ! Our next Meetup is scheduled on Wednesday October 11 at 7:00 pm. Charles Ollion (Heuritech) and Eliot Andres (Freelance) - Deep Learning in production Abstract: We have a slightly different format this time, as we will talk about putting deep learning systems into production. Charles Ollion and Eliot Andres will talk respectively about Training and Inference of Deep Convolutionnal Neural networks in enterprise and production environment. ------------------------------------ Nicolas Patry (Kwyk) - Word2Vec at school Abstract: Kwyk shows you how they use Deep Learning to estimate the level of high school students. ------------------------------------ Balazs Kégl (LAL) - Detecting Mars craters in satellite images Abstract: Balazs will present the new RAMP challenge that focuses on detecting Mars craters. ------------------------------------ Food and drinks will be provided. Heuritech Meetup Team.

    3
  • Deep Learning Meetup #10

    Location visible to members

    Dear Members, Wednesday 07 the 10th Deep learning Meetup will take place at Heuritech! Speakers: Grégory Châtel (Intel) Adversarial examples in deep learning Deep learning is used in more critical systems every day, from self-driving cars to video surveillance. The purpose of this talk is to show how such systems can be fooled using specifically crafted inputs and how to try to defend these kinds of attack Johan Ferret (DreamQuark) Siamese Architectures for Question Answering Mastering Question Answering is a mandatory step towards building efficient conversational interfaces and machine comprehension for corpora of complex documents such as law documents or patient medical records. We describe a supervised approach using Siamese Networks and attention mechanisms to answer non-factoid questions related to the insurance sector. Hedi Ben Younes (LIP6 / Heuritech) Multimodal Tucker Fusion for Visual Question Answering Bilinear models provide an appealing framework for mixing and merging information in Visual Question Answering (VQA) tasks. We introduce MUTAN, a multimodal tensor-based Tucker decomposition to efficiently parametrize bilinear interactions between visual and textual representations. With MUTAN, we control the complexity of the merging scheme while keeping nice interpretable fusion relations. We show how our MUTAN scheme generalizes some of the latest VQA architectures, providing state-of-the-art results. Again, feel free to contact me if you'd like to be a speaker for a next meetup. Intel Sponsorship Intel also created a mini contest for all the participants! If you have a ML project and want to showcase it, share it, or collaborate it with others, submit it to DevMesh (https://devmesh.intel.com/groups/447). Everyone who submits a project to DevMesh will get remote access to Machine Learning Servers. On top of that, best projects will be selected and each winner will receive a gift card (amount TBA). Instructions to join DevMesh: 1/ Create a new account at devmesh.intel.com (https://software.intel.com/registration/dev-mesh/?TARGET=HTTPS%3a%2f%2fsfederation%2eintel%2ecom%2ffederation%2fIDZ_Devmesh%2easp%3fSAMLRequest%3dfZJNT8MwDIb%252FSk%252Fk1PWDorGwDlVUSJMAofFx2AW5ibdFpEmJUwb8erIC2jiAFCmS%252Ffp9bCdTglZ3vOr9xizwpUfyUUWEzitrLqyhvkV3h%252B5VCXxYXJVs431HPEkkvrZIm5EyHvVI2DaB4JHs7BIBWjcgns81mHUPayyPoO3OtuoDnCytaWy4lVmzqA48ZWAH21vTCiW6IXhgvw8m83r5VH%252FzgToWzeuSPRXN6gRPiyaeZCjjYjLO44kUq3gsGimzosjS4yZIiXqcG%252FJgfMnyNBvHaRHOfZ7xIuX5yZJFj%252Bho6CgfpSx6a7UhvhusZL0z3AIp4gYCnnvB76rrKx6EHH7WdljS%252FV%252FTOeutsJrNpjs1H7pzs%252B%252FlxiEtp8lhZvr1XjfBaV7fWq3Ee1RpbbcXDsFjybzrkUWX1rXg%252F2Zno2yIKBmvBinHFpSupHRIxJLZF%252FX3x5h9Ag%253D%253D&AppID=518) 2/ Join your dedicated group - Artificial Intelligence Europe (https://devmesh.intel.com/groups/447) 3/ To submit a project, click on "add a project" *when submitting your project, make sure to select "Artificial Intelligence Europe" as your group. To receive invitations for Intel Machine Learning and Deep Learning webinars, news and tools register in the link below: http://intel.ly/2rVizev (http://intel.ly/2n8pjV7) See you next week, Charles

    9
  • Deep Learning Meetup #9

    Heuritech

    Dear DeepLearners, Our next Meetup is scheduled on Wednesday April 19 at 7:00 pm, with 4 (amazing) speakers : • Alexis Conneau (Facebook AI Research): Transfer Learning for Sentence Embeddings We present a new approach for learning high-quality sentence embeddings. In Computer Vision, image embeddings are usually obtained using ConvNets trained on large supervised datasets. For natural language processing, most successful approaches for word or sentence embeddings are based on architectures trained on large unsupervised corpus. Inspired by the analogy with computer vision, we describe an approach that leverages high-quality supervision from natural language inference dataset and demonstrates superior performance to previous supervised and unsupervised approaches • Rahma Chaabouni (ENS): Multimodal Siamese Networks Recent works have explored deep architectures for learning multimodal speech representation (e.g. audio and images, articulation and audio) in a supervised way. Here we investigate the role of combining different speech modalities, i.e. audio and visual information representing the lips’ movements, in a weakly supervised way using Siamese networks and lexical same-different side information. In particular, we ask whether one modality can benefit from the other to provide a richer representation for phone recognition in a weakly supervised setting. Furthermore, we present a qualitative analysis of the obtained phone embeddings, and show that cross-modal visual input can improve the discriminability of phonological features which are visually discernable (rounding, open/close, labial place of articulation), resulting in representations that are closer to abstract linguistic features than those based on audio only • Loris Felardos (Heuritech): Generative Adversarial Networks and Variational Autoencoders GANs and VAEs produce state of the art results in a number of tasks like image generation and semi-supervised learning. The purpose of this talk is to demonstrate how this is achieved from an engineering point of view, what is the mathematical framework behind those models, and how they can be merged together. More information available on this blog post: https://blog.heuritech.com/2017/04/11/began-state-of-the-art-generation-of-faces-with-generative-adversarial-networks/ Update: Lightning Talk Microsoft Cognitive Toolkit (aka CNTK) & Project Malmo • Morgan Funtowicz (Software Development Engineer, Microsoft Research Cambridge) In this 5’ lightning talk, Morgan Funtowicz will cover the most recent enhancements to the “Project Malmo” (*) platform through an integration with Cognitive Toolkit (https://www.microsoft.com/en-us/research/product/cognitive-toolkit/), the deep learning library of Microsoft. (*) Project Malmo (https://github.com/Microsoft/malmo) is an open source platform for Artificial Intelligence experimentation and research built on top of Minecraft. Food and drinks will be provided. Alexandre

    11
  • Deep Learning Meetup #8 : RNN's with TensorFlow

    In this special edition, we will welcome Martin Görner from Google for a tutorial on recurrent neural networks and Batch normalization with TensorFlow! This presentation is destined to beginners with basic prior knowledge of TensorFlow. It will last approximately one hour. We will provide drinks and snacks after Martin's presentation. Abstract Deep learning has already revolutionized machine learning research, but it hasn't been broadly accessible to many developers. In this session, we'll explore the possibilities of recurrent neural networks by building a language model in TensorFlow. What this model can do will impress you. We welcome developers with no prior machine learning experience. We do recommend that you watch the session "TensorFlow and deep learning without a PhD part 1 (https://youtu.be/qyvlt7kiQoI)" unless you already know about dense and convolutional networks and are only interested in recurrent networks. This session is an intense technical session designed to help beginners in machine learning ramp up quickly. Looking forward to meeting you again!

    6