Publications

2021

• Side channel attacks for architecture extraction of neural networks
  
  • Chabanne Hervé
  • Danger Jean‐luc
  • Guiga Linda
  • Kühne Ulrich
  CAAI Transactions on Intelligent Technologies, Institution of Engineering and Technology, 2021, 6 (1), pp.3-16. (10.1049/cit2.12026)
  DOI : 10.1049/cit2.12026
• Ultra-Low Power system for atrioventricular synchronization using leadless pacemakers
  
  • Maldari Mirko
  • Jabbour Chadi
  • Haddab Youcef
  • Desgreys Patricia
  Bulletin of the International Union of Radio Science, 2021 (376).
• Generalized Mixed-Criticality Static Scheduling for Periodic Directed Acyclic Graphs on Multi-Core Processors
  
  • Medina Roberto
  • Borde Etienne
  • Pautet Laurent
  IEEE Transactions on Computers, Institute of Electrical and Electronics Engineers, 2021, 70 (3), pp.457-470. In safety-critical systems many software components of different criticalities or assurance levels need to interact in a timely manner to keep the system and environment safe. Nowadays, these systems are challenged by technological progress resulting in rapid increases in both software complexity and processing demands. Efficiently designing safety-critical systems subject to stringent timing requirements is therefore a challenge and a necessity. In this article, we consider the mixed-criticality execution model and homogeneous multi-core processors. We begin by defining a task model incorporating mixed-criticality, real-time and precedence constraints in the form of directed acyclic graphs. A meta-heuristic to solve the scheduling problem of this task model is then defined and proved to respect deadlines, even when the system needs to give more processing power to the most critical tasks. The state-of-the-art techniques capable of scheduling a similar task model have only been developed for dual-criticality systems. Conversely, the meta-heuristic we propose has been generalized to support an arbitrary number of criticality levels. We instantiated our meta-heuristic adopting scheduling algorithms such as G-EDF, G-LLF, or G-EDZL for each level of criticality. The experiments show excellent results in terms of acceptance ratio and number of preemptions. (10.1109/TC.2020.2990229)
  DOI : 10.1109/TC.2020.2990229
• NFV Platforms: Taxonomy, Design Choices and Future Challenges
  
  • Zhang Tianzhu
  • Qiu Han
  • Linguaglossa Leonardo
  • Cerroni Walter
  • Giaccone Paolo
  IEEE Transactions on Network and Service Management, IEEE, 2021, 18 (1), pp.30-48. Due to the intrinsically inefficient service provisioning in traditional networks, Network Function Virtualization (NFV) keeps gaining attention from both industry and academia. By replacing the purpose-built, expensive, proprietary network equipment with software network functions consolidated on commodity hardware, NFV envisions a shift towards a more agile and open service provisioning paradigm. During the last few years, a large number of NFV platforms have been implemented to facilitate the development, deployment, and management of Virtual Network Functions (VNFs). Nonetheless, just like any complex system, such platforms commonly consist of abounding software and hardware components and usually incorporate disparate design choices based on distinct motivations or use cases. This broad collection of convoluted alternatives makes it extremely arduous for network operators to make proper choices. Although numerous efforts have been devoted to investigating different aspects of NFV, none of them specifically focused on NFV platforms or attempted to explore their design space. In this paper, we present a comprehensive survey on the NFV platform design. Our study solely targets existing NFV platform implementations. We begin with a top-down architectural view of the standard reference NFV platform and present our taxonomy of existing NFV platforms based on what features they provide in terms of a typical network function life cycle. Then we thoroughly explore the design space and elaborate on the implementation choices each platform opts for. We also envision future challenges for NFV platform design in the incoming 5G era. We believe that our study gives a detailed guideline for network operators or service providers to choose the most appropriate NFV platform based on their respective requirements. Our work also provides guidelines for implementing new NFV platforms. (10.1109/TNSM.2020.3045381)
  DOI : 10.1109/TNSM.2020.3045381
• Linear codes with one-dimensional hull associated with Gaussian sums
  
  • Qian Liqin
  • Cao Xiwang
  • Mesnager Sihem
  Cryptography and Communications - Discrete Structures, Boolean Functions and Sequences, Springer, 2021, 13 (2), pp.225-243. (10.1007/s12095-020-00462-y)
  DOI : 10.1007/s12095-020-00462-y
• A direct proof of APN-ness of the Kasami functions
  
  • Carlet Claude
  • Kim Kwang Ho
  • Mesnager Sihem
  Designs, Codes and Cryptography, Springer Verlag, 2021, 89 (3), pp.441-446. (10.1007/s10623-020-00830-y)
  DOI : 10.1007/s10623-020-00830-y
• Locate UWB Smart Keys
  
  • Park Jiwoong
  • Choi Hong-Beom
  • Ko Young-Bae
  • Lim Keun-Woo
  , 2021, pp.169-171. This poster proposes an Ultra-wideband localization scheme for the smart key system based on a machine learning approach. Previous studies try to use the channel impulse response measurement and apply machine learning to improve the localization accuracy. However, it was found difficult to perform real-time localization in the key fob hardware as these studies require complex computation. In addition, localization accuracy is severely degraded due to frequent NLOS conditions in vehicular environments. In order to solve these problems, we propose vehicular features based on machine learning approach to improve localization accuracy and to support the real-time operation. We evaluate the proposed scheme by deploying Decawave UWB transceivers on an actual vehicle. The proposed scheme showed more than 98% localization accuracy and an average prediction delay of 7.4 milliseconds. (10.1145/3446382.3448884)
  DOI : 10.1145/3446382.3448884
• Minage de règles rapide, exact et exhaustif dans de larges bases de connaissances
  
  • Lajus Jonathan
  , 2021. The Semantic Web has quickly become a constellation of large and interconnected entity-centric Knowledge Bases. These KBs contain domain-specific knowledge that can be used for multiple application such as question answering or automatic reasoning. But in order to take full advantage of this data, it is essential to understand the schema and the patterns of the KB. A simple and expressive manner to describe the dependencies in a KB is to use rules. Thus it is crucial to be able to perform rule mining at scale.In this thesis, we introduce novel approaches and optimizations designed to speed up the process of rule mining on large Knowledge Bases. We present two algorithms that implements these optimizations: the AMIE 3 algorithm (the successor of the exact rule mining algorithm AMIE+) and the Pathfinder algorithm, a novel algorithm specialized in mining path rules. These two algorithms are exhaustive with regard to the parameters provided by the user, they compute the quality measures of each rule exactly and efficiently scale to large KB and longer rules.
• Recurring concept memory management in data streams: exploiting data stream concept evolution to improve performance and transparency
  
  • Halstead Ben
  • Koh Yun Sing
  • Riddle Patricia
  • Pears Russel
  • Pechenizkiy Mykola
  • Bifet Albert
  Data Mining and Knowledge Discovery, Springer, 2021, 35 (3), pp.796--836. A data stream is a sequence of observations produced by a generating process which may evolve over time. In such a time-varying stream the relationship between input features and labels, or concepts, can change. Adapting to changes in concept is most often done by destroying and incrementally rebuilding the current classifier. Many systems additionally store and reuse previously built models to more efficiently adapt when stream conditions drift to a previously seen state. Reusing a model offers increased classification performance over rebuilding, and provides an indicator, or transparency, into the hidden state of the generating process. When only a subset of past models can be stored for reuse, for example due to memory constraints, the choice of which models to store for optimal future reuse is an important problem. Current methods of evaluating which models to store use valuation policies such as age, time since last use, accuracy and diversity. These policies are often not optimal, losing predictive performance by undervaluing complex models. We propose a new valuation policy based on advantage, the misclassifications avoided by reusing a model rather than training a new model, which more accurately reflects the true value of model storage. We evaluate our method on synthetic and real world data, including a real world air pollution dataset. Our results show accuracy increases of up to 6% using our valuation policy, while preserving transparency. (10.1007/S10618-021-00736-W)
  DOI : 10.1007/S10618-021-00736-W
• Implementing Secure Applications thanks to an Integrated Secure Element
  
  • Guilley Sylvain
  • Le Rolland Michel
  • Quenson Damien
  , 2020, pp.566-571. More and more networked applications require security, with keys managed at the end-point. However, traditional Secure Elements have not been designed to be connected. There is thus a need to bridge the gap, and novel kinds of Secure Elements have been introduced in this respect. Connectivity has made it possible for a single chip to implement multiple usages. For instance, in a smartphone, security is about preventing the device from being rooted, but also about enabling user's online privacy. Therefore, Secure Elements shall be compatible with multiple requirements for various vertical markets (e.g., payment, contents protection, automotive, etc.). The solution to this versatility is the integration of the Secure Element within the device main chip. Such approach, termed iSE (integrated Secure Element), consists in the implementation of a subsystem, endowed to manage the chip security, within a host System-on-Chip. The iSE offers flexibility in the security deployment. However, natural questions that arise are: how to program security applications using an iSE? How to certify those applications, most likely according to several different schemes? This position paper addresses those questions, and comes up with some key concepts of on-chip security, in terms of iSE secure usage. In particular, we will show in this paper that iSE nowadays shall be designed so that the product it embeds is certifiable in a multiplicity of schemes, and so even before the product is launched on the market. (10.5220/0010298205660571)
  DOI : 10.5220/0010298205660571
• Interfacing Digital and Analog Models for Fast Simulation and Virtual Prototyping
  
  • Genius Daniela
  • Apvrille Ludovic
  , 2021, pp.224-231. The paper presents an enhancement for the virtual simulation of analog/mixed-signal systems from high-level SysML models. Embedded systems, e.g. in robotic systems, feature both digital and analog circuits such as sensors and actuators. Simulation of these system requires to handle both domains (digital, analog); our aim is thus to make the interactions between these two domains explicit. For this, the paper first defines new send and receive procedures, then explains how to check semantic aspects of models to ensure a correct simulation. A running example illustrates the basic concepts of our approach. A proof of concept based on an existing rover is also presented. (10.5220/0010257202240231)
  DOI : 10.5220/0010257202240231
• Direct Model-checking of SysML Models
  
  • Tempia Calvino Alessandro
  • Apvrille Ludovic
  , 2021, pp.216-223. Model-checking intends to verify whether a property is satisfied by a model, or not. Model-checking of high-level models, e.g. SysML models, usually first requires a model transformation to a low level formal specification. The present papers proposes a new model-checker that can be applied (almost) directly to the SysML model. The paper first explains how this model-checker works. Then, we explain how it can efficiently check CTL-like properties. Finally, the paper discusses the performance of this model-checker integrated in the TTool framework. (10.5220/0010256302160223)
  DOI : 10.5220/0010256302160223
• Almost sure convergence rates for Stochastic Gradient Descent and Stochastic Heavy Ball
  
  • Sebbouh Othmane
  • Gower Robert M
  • Defazio Aaron
  , 2021. We study stochastic gradient descent (SGD) and the stochastic heavy ball method (SHB, otherwise known as the momentum method) for the general stochastic approximation problem. For SGD, in the convex and smooth setting, we provide the first almost sure asymptotic convergence rates for a weighted average of the iterates. More precisely, we show that the convergence rate of the function values is arbitrarily close to o(1/ √ k), and is exactly o(1/k) in the so-called overparametrized case. We show that these results still hold when using stochastic line search and stochastic Polyak stepsizes, thereby giving the first proof of convergence of these methods in the non-overparametrized regime. Using a substantially different analysis, we show that these rates hold for SHB as well, but at the last iterate. This distinction is important because it is the last iterate of SGD and SHB which is used in practice. We also show that the last iterate of SHB converges to a minimizer almost surely. Additionally, we prove that the function values of the deterministic HB converge at a o(1/k) rate, which is faster than the previously known O(1/k). Finally, in the nonconvex setting, we prove similar rates on the lowest gradient norm along the trajectory of SGD.
• Self-secured PUF: Protecting the Loop PUF by Masking
  
  • Tebelmann Lars
  • Danger Jean-Luc
  • Pehl Michael
  , 2021, pp.293-314. Physical Unclonable Functions (PUFs) provide means to gen-erate chip individual keys, especially for low-cost applications such as theInternet of Things (IoT). They are intrinsically robust against reverseengineering, and more cost-effective than non-volatile memory (NVM).For several PUF primitives, countermeasures have been proposed to mit-igate side-channel weaknesses. However, most mitigation techniques re-quire substantial design effort and/or complexity overhead, which can-not be tolerated in low-cost IoT scenarios. In this paper, we first ana-lyze side-channel vulnerabilities of the Loop PUF, an area efficient PUFimplementation with a configurable delay path based on a single ringoscillator (RO). We provide side-channel analysis (SCA) results frompower and electromagnetic measurements. We confirm that oscillationfrequencies are easily observable and distinguishable, breaking the se-curity of unprotected Loop PUF implementations. Second, we presenta low-cost countermeasure based on temporal masking to thwart SCAthat requires only one bit of randomness per PUF response bit. The ran-domness is extracted from the PUF itself creating aself-secured PUF.The concept is highly effective regarding security, low complexity, andlow design constraints making it ideal for applications like IoT. Finally,we discuss trade-offs of side-channel resistance, reliability, and latency aswell as the transfer of the countermeasure to other RO-based PUFs (10.1007/978-3-030-68773-1_14)
  DOI : 10.1007/978-3-030-68773-1_14
• Analyse automatique des comportements multimodaux lors d’entretiens vidéo différés pour le recrutement
  
  • Hemamou Léo
  , 2021. Le développement des nouvelles technologies influence tous les secteurs d’activités, y compris celui des ressources humaines et notamment le processus de recrutement. L’émergence des entretiens vidéo différés permet d’organiser en asynchrone des entretiens avec des candidats et de les évaluer. Les candidats se connectent à une plateforme et se filment pendant qu’ils répondent à des questions définies en amont par les recruteurs. La plateforme permet ensuite à plusieurs recruteurs d’évaluer le candidat, de partager des notes et d’inviter éventuellement le candidat à un entretien en face-à-face. Le nombre de telles candidatures vidéo devient cependant de plus en plus volumineux et difficile à traiter « manuellement » par un ou deux recruteurs. Il devient donc nécessaire d’envisager une aide pour le recruteur qui doit traiter parfois plusieurs centaines d’entretiens vidéo. De plus, le développement d'une telle aide pourra aussi permettre aux candidats de s'entraîner à l'exercice de l'entretien vidéo différé grâce à une évaluation automatique. Dans le cadre d'un projet avec un partenaire industriel, nous avons recueilli deux corpus de plus de 5000 entretiens d'embauche vidéo asynchrones pour des postes réels. Cette thèse étudie la tâche consistant à prédire les performances des candidats lors d'entretiens vidéo asynchrones en utilisant trois modalités (contenu verbal, prosodie et expressions faciales) grâce à des données provenant d'entretiens réels qui se déroulent dans des conditions hors laboratoires. Nous proposons un nouveau modèle multimodal d'attention hiérarchique appelé HireNet qui vise à prédire l’employabilité des candidats tels qu'ils sont évalués par les recruteurs. Dans HireNet, un entretien est considéré comme une séquence de questions et de réponses contenant des signaux sociaux saillants. Dans un second temps, nous abordons la question de l’influence du comportement non verbal dans une décision d'embauche. Il est important d'étudier cette question car elle pourrait permettre de mieux comprendre comment former les candidats aux entretiens d'embauche et sensibiliser les recruteurs à ces comportements influents. Par la suite, nous nous concentrons sur l'étude des signaux sociaux influents dans les entretiens vidéo d'embauche asynchrones qui sont découverts par les méthodes d'apprentissage profond. Une particularité de HireNet est l'utilisation de mécanismes d'attention qui visent à identifier les parties les plus pertinentes d'une réponse. Ainsi, des informations à un niveau temporel fin pourraient être extraites en utilisant des annotations globales (au niveau de l'entretien) sur la convocabilité du candidat. Alors que la plupart des systèmes d'apprentissage profond utilisent des mécanismes d'attention pour offrir une visualisation rapide des tranches lorsqu'une augmentation des valeurs d’attention se produit, nous effectuons une analyse approfondie pour comprendre ce qui se passe lors de ces moments. Ainsi, nous étudions le contenu de ces moments en les comparant avec des moments échantillonnés de manière aléatoire afin d’étudier leurs différences. Dans l’ensemble, cette méthode vise à améliorer l’interprétabilité de tels systèmes et à s’interroger sur leur utilisation comme outil exploratoire. Notre troisième contribution concerne les biais dans les systèmes d'analyse automatique des entretiens vidéo. Nous proposons une première approche qui utilise un entraînement adversaire pour apprendre une représentation ignorant le sexe et l'ethnicité des candidats. Nous montrons expérimentalement qu'elle assure une meilleure représentation sans perte significative d'efficacité sur la tâche principale. Nous étudions ensuite l'utilisation de cet entraînement adversaire sans qu'il soit nécessaire de recueillir des informations sensibles sur les candidats. Ainsi, nous visons à améliorer l'équité des prochains systèmes automatiques de traitement des vidéos d'entretiens d'embauche pour une égalité dans la sélection des emplois.
• A lightweight neural model for biomedical entity linking
  
  • Chen Lihu
  • Varoquaux Gaël
  • Suchanek Fabian
  , 2021, 35 (14), pp.12657-12665. Biomedical entity linking aims to map biomedical mentions, such as diseases and drugs, to standard entities in a given knowledge base. The specific challenge in this context is that the same biomedical entity can have a wide range of names, including synonyms, morphological variations, and names with different word orderings. Recently, BERT-based methods have advanced the state-of-the-art by allowing for rich representations of word sequences. However, they often have hundreds of millions of parameters and require heavy computing resources, which limits their applications in resource-limited scenarios. Here, we propose a lightweight neural method for biomedical entity linking, which needs just a fraction of the parameters of a BERT model and much less computing resources. Our method uses a simple alignment layer with attention mechanisms to capture the variations between mention and entity names. Yet, we show that our model is competitive with previous work on standard evaluation benchmarks.
• Triple Sensing Current Margin for Maintainable MRAM Yield at Sub-100% Tunnel Magnetoresistance Ratio
  
  • Cai Hao
  • Han Menglin
  • Zhou Yongliang
  • Liu Bo
  • Naviner Lirida
  IEEE Transactions on Magnetics, Institute of Electrical and Electronics Engineers, 2021, 57 (2), pp.1-5. Spin transfer torque magnetic random access memory (STT-MRAM) creates significant breakthroughs as a proper candidate of next-generation non-volatile memory (NVM). Although STT-MRAM achieves high endurance, low access latency, and power consumption, the yield issue remains one of the critical concerns in high-density and large-scale MRAM array design. In this article, a novel triple-current margin sensing amplifier (TM-SA) is proposed for maintainable MRAM yield based on the current-mode SA and transmission gate switches. The sensing current margin of the proposed TM-SA is three times enlarged compared to traditional current mean (CM)-SA and resistance mean (RM)-SA. With a seriously degraded tunnel magnetoresistance (TMR) ratio (sub-100%, as low as 10%), the maximum voltage margin is 4.6 times of conventional CM-SA and five times of RM-SA. Monte-Carlo simulation shows that sensing failure probability can be greatly alleviated with the proposed TM-SA. The performance of TM-SA with respect to voltage margin can be further improved than that of CM-SA and RM-SA. (10.1109/TMAG.2020.3011614)
  DOI : 10.1109/TMAG.2020.3011614
• Making Obfuscated PUFs Secure Against Power Side-Channel Based Modeling Attacks
  
  • Danger Jean-Luc
  • Kroeger Trevor
  • Cheng Wei
  • Guilley Sylvain
  • Karimi Nazhmeh
  , 2021, pp.1000-1005. (10.23919/DATE51398.2021.9474137)
  DOI : 10.23919/DATE51398.2021.9474137
• Ombroscopic imaging and PTV to reduce the depth of field in a flow with strong curvature effects
  
  • Fokoua Georges
  • Durand Antoine
  • Cuvelier Philippe
  Optics and Lasers in Engineering, Elsevier, 2021, 137, pp.106386. The present work deals with the depth of field characterization in a complex flow with strong curvature effects: the bubbly Taylor Couette flow. Measurements were carried out in a 2D (axial and radial) plane with a large contribution of the azimuthal velocity on radial and axial velocity. Ombroscopic imaging technique coupled with a Particle Tracking Velocimetry have been used to accurately measure the depth of field in the working zone. We show that reducing this depth of field helps to minimize dispersed phase measurement errors in the radial direction for both position and velocity. We also found good agreement between results of the proposed method and the theoretical predictions of bubbles velocity in laminar Taylor-Couette flow. (10.1016/j.optlaseng.2020.106386)
  DOI : 10.1016/j.optlaseng.2020.106386
• Enhancement of sensing range of Brillouin optical time‐domain reflectometry system up to 150 km with in‐line bi‐directional erbium‐doped fibre amplifications
  
  • Clement Pierre
  • Gabet Renaud
  • Lanticq Vincent
  • Jaouën Yves
  Electronics Letters, IET, 2021, 57 (3), pp.142-144. (10.1049/ell2.12012)
  DOI : 10.1049/ell2.12012
• On the tensor rank of multiplication in finite extensions of finite fields and related issues in algebraic geometry
  
  • Ballet Stéphane
  • Pieltant Julia
  • Chaumine Jean
  • Rambaud Matthieu
  • Randriambololona Hugues
  • Robert Rolland
  Russian Mathematical Surveys, Turpion, 2021, 76 (1), pp.29-89. In this paper, we give a survey of the known results concerning the tensor rank of multiplication in finite extensions of finite fields, enriched with some unpublished recent results, and we analyze these to enhance the qualitative understanding of the research area. In particular, we identify and clarify certain partially proved results and emphasise links with open problems in number theory, algebraic geometry, and coding theory. Bibliography: 92 titles. (10.1070/RM9928)
  DOI : 10.1070/RM9928
• Beamwidth optimization and resource partitioning scheme for localization assisted mm-wave communications
  
  • Ghatak G.
  • Koirala R.
  • de Domenico A.
  • Denis B.
  • Dardari D.
  • Uguen Bernard
  • Coupechoux Marceau
  IEEE Transactions on Communications, Institute of Electrical and Electronics Engineers, 2021, 69 (2), pp.1358-1374. We study a millimeter wave (mm-wave) wireless network deployed along the roads of an urban area, to support localization and communication services simultaneously for outdoor mobile users. In this network, we propose a mm-wave initial beam-selection scheme based on localization-bounds, which greatly reduces the initial access delay as compared to traditional initial access schemes for standalone mm-wave small cell base station (BS). Then, we introduce a downlink transmission protocol, in which the radio frames are partitioned into three phases, namely, initial access, data, and localization, respectively. We establish a trade-off between the localization and communication performance of mm-wave systems, and show how enhanced localization can actually improve the data-communication performance. Our results suggest that dense BS deployments enable to allocate more resources to the data phase while still maintaining appreciable localization performance. Furthermore, for the case of sparse deployments and large beam dictionary size (i.e., with thinner beams), more resources must be allotted to the localization phase for optimizing the rate coverage. Based on our results, we provide several system design insights and dimensioning rules for the network operators that will deploy the first generation of mm-wave BSs. (10.1109/TCOMM.2020.3036864)
  DOI : 10.1109/TCOMM.2020.3036864
• Radiofrequency electromagnetic fields from mobile communication: Description of modeled dose in brain regions and the body in European children and adolescents
  
  • Birks Laura Ellen
  • van Wel Luuk
  • Liorni Ilaria
  • Pierotti Livia
  • Guxens Mònica
  • Huss Anke
  • Foerster Milena
  • Capstick Myles
  • Eeftens Marloes
  • El Marroun Hanan
  • Estarlich Marisa
  • Gallastegi Mara
  • Safont Llúcia González
  • Joseph Wout
  • Santa-Marina Loreto
  • Thielens Arno
  • Torrent Maties
  • Vrijkotte Tanja
  • Wiart Joe
  • Röösli Martin
  • Cardis Elisabeth
  • Vermeulen Roel
  • Vrijheid Martine
  Environmental Research, Elsevier, 2021, 193, pp.110505. (10.1016/j.envres.2020.110505)
  DOI : 10.1016/j.envres.2020.110505
• Handling causality and schedulability when designing and prototyping cyber-physical systems
  
  • Cortés porto Rodrigo
  • Genius Daniela
  • Apvrille Ludovic
  Software and Systems Modeling, Springer Verlag, 2021, 20 (1). Cyber physical systems are built upon digital and analog circuits, making it necessary to handle different models of computation during their design and verification (e.g., by simulation). When designing these systems, an important aspect to consider is the causality between the different domains. For this, we introduce a new model-driven framework able to identify causality problems and to suggest a valid schedule between the analog and digital domains. Once a valid schedule has been computed, our framework can generate cycle and bit accurate virtual prototypes (in SystemC/SystemC AMS) from high-level SysML models. (10.1007/s10270-021-00866-1)
  DOI : 10.1007/s10270-021-00866-1
• Sparse Realization in Unreliable Spin-Transfer-Torque RAM for Convolutional Neural Network
  
  • Cai Hao
  • Chen Juntong
  • Zhou Yongliang
  • Hong Xiaofeng
  • Liu Bo
  • Naviner Lirida
  IEEE Transactions on Magnetics, Institute of Electrical and Electronics Engineers, 2021, 57 (2), pp.1-5. The explosive growth of in-memory computing and neural network requires stringent demands on the computational energy efficiency. Nonvolatile memories such as magnetic random access memory (MRAM) provides alternative memory solutions toward energy efficiency. Sparsity realization across emerging device, hybrid circuit, and algorithmic becomes a recent trend in neural network. Previous sparse adaption in memories mainly focused on high level analysis. In this article, the sparse realization of hybrid magnetic/CMOS integration is first proposed for convolutional neural network (CNN). Simulation results with representative data sets CIFAR-10 show that MRAM sensing operation can be speedup 6.4× with 84.46% sparsity. The proposed training and retraining phases can solve unreliable sensing issues with a proper sparsity selection. (10.1109/TMAG.2020.3015146)
  DOI : 10.1109/TMAG.2020.3015146