SOMA Research Project About Impact Partners Results ☰
About Impact Partners Results

SOMA

neural-based Self-Organizing Machine Architecture

A research project funded by French (ANR) and Swiss (FNS) research agencies.

AboutImage

About SOMA research project

The tremendous increase of transistors integration during the last few years has reached the limits of classic Von Neuman architectures. This has led to a wide adoption of parallel processors by the industry, enabling many-core processing architectures as a natural trend for the next generation of computing devices. Nonetheless, one major issue of such massively parallel processors is the design and the deployment of applications that cannot make an optimal use of the available hardware resources.
This limit is even more acute when we consider application domains where the system evolves under unknown and uncertain conditions such as mobile robotics, IoT, autonomous vehicles or drones.
In the end, it is impossible to foresee every possible context that the system will face during its lifetime, making thus impossible to identify the optimal hardware substrate to be used.

The biological answer

Interestingly enough, the biological brain has ”solved” this problem using a dedicated architecture and mechanisms that offer both adaptive and dynamic computations, namely, self-organization. However, even if neuro-biological systems have often been a source of inspiration for computer science (as recently demonstrated by the renewed interest in deep-learning), the transcription of self-organization at the hardware level is not straightforward and requires a number of challenges to be taken-up.

  • The first challenge is to extend the usual self-organization mechanisms to account for the dual levels of computation and communication in a hardware neuromorphic architecture. From a biological point of view, this corresponds to a combination of the so-called synaptic and structural plasticities. We intend to define computational models able to simultaneously self-organize at both levels, and we want these models to be hardware-compliant, fault tolerant and scalable by means of a neuro-cellular structure.
  • The second challenge is to prove the feasibility of a self-organizing hardware structure. Considering that these properties emerge from large scale and fully connected neural maps, we will focus on the definition of a self-organizing hardware architecture based on digital spiking neurons that offer hardware efficiency.
  • The third challenge consists in coupling this new computation paradigm with an underlying conventional manycore architecture. This will require the specification of a Network-on-Chip that adapts to self-organizing hardware resources, as well as the definition of a programming model using the learning of input data to better and automatically divide and allocate functional elements.

BioBlockImage
MultiDispImage

A multidisciplinary project.

From multiple expertises.

Hence, this project is a convergence point between past research approaches toward new computation paradigms: adaptive reconfigurable architecture, cellular computing, computational neuroscience, and neuromorphic hardware.

  • 1. SOMA is an adaptive reconfigurable architecture to the extent that it will dynamically re-organize both its computation and its communication by adapting itself to the data to process.
  • 2. SOMA is based on cellular computing since it targets a massively parallel, distributed and decentralized neuromorphic architecture.
  • 3. SOMA is based on computational neuroscience since its selforganization capabilities are inspired from neural mechanisms.
  • 4. SOMA is a neuromorphic hardware system since its organization emerges from the interactions between neural maps transposed into hardware from brain observation.

A significant step

This project represents a significant step toward the definition of a true fine-grained distributed, adaptive and decentralized neural computation framework. Using self-organized neural populations onto a cellular machine where local routing resources are not separated from computational resources, it will ensure natural scalability and adaptability as well as a better performance/power consumption tradeoff compared to other conventional embedded solutions. This new computing framework may indeed represent a viable integration of neuromorphic computing into the classical Von Neumann architecture and could endow these hardware systems with novel adaptive properties.

SigStepImage
MainGoalsImage

Main goals

Five main goals

  • Follow a bio-inspired approach
  • Study brain plasticity
  • Define new neural models
  • Design neuromorphic hardware
  • Develop a FPGA-based platform for brain-inspired processing

×

Credits

Funding

Images

Web edition

Web site by Laurent Rodriguez powered by LBO (LTech)

SOMA

neural-based Self-Organizing Machine Architecture

A research project funded by French (ANR) and Swiss (FNS) research agencies.

Impact and benefits of the project

Scientific impact

Neuromorphic computing becomes a subject of research and innovation that is becoming increasingly important at the European and international level. The report of the OMNT (CEA / CNRS) identifies this research topic as strategic by 2025 and shows the explosion already in progress in the fields of application using the results of this research. In addition, the OMNT working group also shows that many research groups and start-up are established in the united states. The power of artificial neural networks has already been widely confirmed with the proliferation of deep networks on many databases of different natures. But the computational and energetic cost of these new algorithms also shows the importance of studying in a coupled way the hardware implantation through neuromorphic architectures. As identified by the CNRS research group BioComp (http://gdr-biocomp.fr/) the next challenges of these architectures relate to the technological aspects of the resistive memories better adapted to the implementation of the synaptic functions and the unsupervised learning rules, and from the architectural point of view the transition to large networks ensuring the population dynamics and their self-organizing properties. This second point will be studied in the project that addresses different scientific challenges at different levels of implementation. The project deals with brain mechanisms modeling at neurobiological level to hardware design and construction of novel computing devices. In between, we will develop the bridge for linking both domains, and end up with an adaptive system endowed with brain-like plasticity capabilities. This project vision implies a new paradigm for building future computing systems. We will end up with a multiprocessing distributed system, without any central coordinator, where connections topologies are incrementally created in a self-organizing manner, and with the capability of self-heal in the presence of failure in connections and/or computing nodes. A final validation on a real working hardware platform represents a major proof of concept in this type of projects. We focus on hardware mechanisms to be implemented on real hardware, which will impose us hard constraints to meet. The implementation-oriented approach will be an important breackthrough in the domain of cellular computing architectures given the unprecedented computing capabilities of the selected devices combined with the novel adaptability and scalability features to be developed in SOMA

Benefits of the project

Progress and breaks in ICT are based, among other things, on improving the performance of devices that process or transfer information. These devices must meet application challenges such as energy efficiency, compactness or resiliency of systems, connected objects, or man-machine collaboration, but also make the emergence of safe systems possible for the exploitation of large data sets (Elements of the National Research Strategy). In his context the Action Plan 2017 of the ANR rises the question of alternative architectures such as bio-inspired computing and neuromorphic architectures. Beyond the short-term scientific interest on building neuromorphic systems for endowing current reconfigurable devices with adaptability features, we have already identified a potential range of mid-term applications for these devices. As part of that review, we have gathered a list of end-user applications that may benefit from the results of this project. The main type of applications identified in here, are those involving systems interacting with dynamic environments. Application fields like wearable computing, IoT, autonomous robotics, domotics, autonomous vehicles and unmanned air vehicles, among others are systems that are constantly evolving in a changing world. They have to deal with imprecision, noisy inputs, faulty sensors, unreliable data. This type of systems will largely benefit from the resulting features offered by our hardware platform.

Dissemination

The project Coordinator will manage the dissemination/exploitation of the scientific and technical results from this project supported by the partner institutions. In addition, individual partners will be responsible for the execution of the actual dissemination/exploitation activities, especially during a special session of the BioComp workshops, organized by the LEAT. Newly acquired knowledge will continuously be evaluated for protection, dissemination and exploitation of intellectual property rights by each partner. In order to address and stimulate the interest of industry, research organizations, including academia in our research we are proposing a number of ways the results of our research will be disseminated: (1) Internet (a website for the project will be created), (2) scientific publications and presentations in journals and conferences, (3) creation of a master course on neural-based adaptive architectures. (In the past we have already established collaboration on teaching activities between the partners from HES-SO and U. Nice), (4) exhibiting our FPGA implemented demonstrator at international events, and (5) organization of a workshop on neuromorphic systems. The Coordinator will be responsible for ensuring that a secure and suitable knowledge management system is put in place, which will run as soon as possible after the project has started.

SOMA

neural-based Self-Organizing Machine Architecture

A research project funded by French (ANR) and Swiss (FNS) research agencies.

Inria Nicolas Rougier is vice-head of the Mnemosyne project which is hosted at the Institute of Neurodegenerative Diseases in Bordeaux. This project is involved in the study and modelling of major cognitive and behavioral functions (eg. attention, recognition, planning, decision) that are known to emerge from adaptive sensorimotor loops involving the external world, the body and the brain. He has recently investigated visual attention in order to understand what are the inner mechanisms of occular saccades and more recently, he has been investigating self-organization of representations within the primary somato-sensory cortex and the influence of attention in the formation and the refinement of these representations. He’s also investigating the mechanisms of action selection through the modeling of the basal ganglia complex with a special emphasis on the motor cortex. The challenge in all these modelling approaches is to understand how a consistent behavior emerges from a purely distributed, numerical and adaptive computing, i.e. without any form of a central supervisor or homonculus.
LEAT LEAT, the research laboratory in Electronics at the Universite Cote d’Azur, is affiliated to CNRS as UMR 7248. LEAT relies on the scientific disciplines of embedded systems design, modeling and optimization of networks of wireless sensors, System-on-Chip and communicating objects. All the research work of the team is grouped according to four axes: The optimization of the energy consumption in the communicating objects, Reconfigurable self-adaptive systems, Reactive and cooperative systems and Behavioral modeling and modeling of radiofrequency systems. Benoît Miramond is leader of the research group on Reconfigurable and Adaptive Systems, specialized in embedded systems design, multi/many-core architectures, reconfigurable computing and bio-inspired architectures. LEAT is member of the organizing committee of the BioComp group of research of the CNRS. BioComp aims at organizing a community of researchers coming from the domains of computer science, neurosciences, robotics and microelectronics in order to study the principles of natural computing and to transpose it into hardware substrates. In this context Benoît Miramond has been the organizer of the NeuroSTIC workshop since 2011The Laboratory of Electronics, Antennas and Telecommunications (LEAT) is a Joint Unit of the University of Nice Sophia Antipolis (member University Côte d'Azur) and CNRS (UMR No. 7248). Its scientific fields are articulated around two main axes: the first concerns antennas, electromagnetism and microwaves and the second focuses on communicating objects, optimization of wireless networks, embedded systems and systems on Chip (SoC). The LEAT regroups 69 members, 31 of whom are permanent and 38 non-permanent (doctoral students, post-docs, etc.). In particular, one of LEAT's areas of research specializes in the design of neuromorphic maps.
Loria Loria is the French acronym for the “Lorraine Research Laboratory in Computer Science and its Applications” and is a research unit (UMR 7503), common to CNRS, the University of Lorraine and INRIA. The Biscuit team is affiliated with CNRS and University of Lorraine (UL). Bernard Girau (Professor UL) is head of the team. He is also member of the organizing committee of the BioComp group of research of the CNRS. His domain of research relates to computational neurosciences and cellular computing. More precisely, the goal is to study the properties and computational capacities that emerge from unconventional computing architectures for the design or control of persistent autonomous systems that interact with complex dynamic environments. Inspired by microscopic (spiking neurons) and mesoscopic (neural populations, dynamic neural fields) aspects of neural computations, his research aims at generating neuro-cellular paradigms of distributed spatial computation and analyzing their properties (e.g. self-organization, robustness, hardware compliance, etc.).
INIT / Hes-so The inIT institute (INstitut d’ingenierie Informatique et des Télécommunications) at the HES-SO consists of three main research axes. The axis of embedded systems drives different projects related to reconfigurable computing, Internet of Things, self-adaptive hardware systems, sensors and antenna design among others. In addition to Prof. Upegui, Fabien Vannel will also be active in this project. Prof. Vannel is the leader of the digital systems group at inIT, HES-SO. He received his engineering degree from ESIEE (France) and a PhD degree in computer science from the EPFL (Switzerland) for his thesis entitled ”Bio-inspired cellular machines: towards a new electronic paper architecture”. Since 2009, he is professor at the University of Applied Sciences of Western Switzerland – hepia, Geneva. where he his doing his main research activities on interconnected embedded systems and high-performance FPGA computation. Prof. A. Upegui and Prof. F. Vannel have a significant experience on building self-adaptive distributed systems featuring fault-tolerance, bio-inspired systems with epigenetic, ontogenetic, and phylogenetic capabilities, and the applications of such systems to real-world complex systems applications.

SOMA

neural-based Self-Organizing Machine Architecture

A research project funded by French (ANR) and Swiss (FNS) research agencies.

Publications

  • AHS 2018 – "Pruning Self-Organizing Maps for Cellular Hardware Architectures", A. Upegui, B. Girau, N. Rougier, F. Vannel, B. Miramond
  • NIPU 2018 – "Neuromorphic hardware as a self-organizing computing system", L. Khacef, B. Girau, N. Rougier, A. Upegui, B. Miramond
  • IPAS 2018 – "A distributed cellular approach of large scale SOM models for hardware implementation", L. Rodriguez, L. Khacef, B. Miramond
  • ICTAI 2018 – "NP-SOM: network programmable self-organizing maps", Y. Bernard, E. Buoy, A. Fois, B. Girau
  • SSCI 2018 – "SCALP: Self-configurable 3-D Cellular Adaptive Platform", Fabien Vannel, Diego Barrientos, Joachim Schmidt, Christian Abegg, Damien Buhlmann and Andres Upegui, Nov 2018