WP 6 2nd year: Real-Time Hardware/Software Simulation Platforms for Adaptive Spiking Neural Networks

Architecture of the 2nd generation simulation platform (enlarge)

The objectives of this work package are the design and exploitation of a simulation platform based on VLSI circuits that simulate the activity of adaptive neural networks, using conductance-based models (Hodgkin-Huxley formalism).

Note : Real-time performance in simulation of spiking neural networks is an important feature for two main reasons:

  1. it guaranties the computational speed, compared to traditional software processing tools (although massive parallel computing facilities do provide high performance, this is achieved at tremendous cost - which strictly limits the extent to which they can be used);
  2. it provides a simplified interactivity with the natural environment, such as sensory inputs.

The simulation platforms we present are designed to be efficient and convenient tools for the exploration of biologically-realistic networks. Simulations which explore the behavior of a neural network in multi-parameter space are time-costly. The platforms intend to help studying the relatively complex paradigms involving plasticity mechanisms.

The research group "Engineering of Neuromorphic Systems" in IMS-Bordeaux is in charge of designing successive generations of custom VLSi circuits (ASICs) and associated simulation platforms. These circuits compute the activity of neural elements using a complex conductance-based model (Hodgkin-Huxley formalism); The models parameters are provided by WP4. The ASICs are integrated on a simulation platform, where they can be organized as a neural network with an adaptive all-to-all connectivity. The platform is optimized to run temporal plasticity algorithms such as STDP. These algorithms are defined in WP5. Simulations run in biological real-time, and the platform will be used by the other partners to run a systematic exploration of the implemented models and plasticity algorithms. Comparative simulations exploiting the WP6 and WP7 platforms will be conducted during year3, to select and optimize the connectivity and plasticity functions used in large scale simulations.

During the second year of the project, the 3 tasks of WP6 were conducted in parallel. In Task2 (models specification and adaptation for the hardware), we exploited the test results from the first generation of ASICs to evaluate the implementability of the new models defined in WP5. We also set up an IP-based design flow to optimize the cells re-use in the hardware neurons on ASICs. In Task 3, we designed the second generation of the simulation platform, using the ASICs designed during year 1. We also started the optimisation of the STDP implementation on FPGA, depending on the neural network connectivity and complexity of dynamics. Finally, in Task 1, we conducted experiments on adaptive networks of RS neurons with Poisson noise inputs. We also prepared an experimental protocol to evaluate the efficacy of conductance-base neurons when computing cortical maps.

  • Task 1 : Experiments using the hardware simulation platform: single neuron validation, 6-neurons adaptive network with correlated Poisson inputs (see Deliverable D17: Report on hardware simulations of single-cell and network benchmarks).
  • Task 2 : Specification of 4 neurons (conductance-based) models for the next ASICs generation (see WP4). Optimization of the ASICs design process by defining a IP-based library of analog modules for semi-automated IC synthesis.
  • Task 3 : Tests and validation of the first generation of ASICs. Fabrication of the simulation platform (second generation). Adaptation of data and models format to the facets database. Evaluation of an optimized architecture for the third generation platform (distributed STDP computation).


  • Q. Zou, Y. Bornat, T. Lévi, and S. Renaud, "Real-time simulations of networks of Hodgkin–Huxley neurons using analog circuits," Neurocomputing, vol. 69, pp. 1137-1140, 2006 (fulltext.pdf)
  • Q. Zou, Y. Bornat, S. Saïghi, T. Lévi, and S. Renaud, "Analog-digital simulations of full conductance-based networks of spiking neurons," Network: Computation in Neural Systems, vol. 17, pp. 211-233, 2006 (fulltext.pdf)
  • T. Lévi, N. Lewis, Y. Bornat, A. Destexhe, and J. Tomas, "An IP-based library for the design of analog Hodgkin-Huxley neurons," in Proceedings of the Systems, Signal and Devices Conference (SSD 2007), Hammamet, Tunisia, 2007 (fulltext.pdf)
  • N. Lewis and S. Renaud, "Spiking neural networks "in silico": from single neurons to large scale networks," in Proceedings of the Systems, Signal and Devices Conference (SSD 2007), Hammamet, Tunisia, 2007 (fulltext.pdf)
  • O. Rochel, S. Chemla, P. Kornprobst, T. Viéville, A. Daouzli, S. Saïghi, C. Lopez, and S. Renaud, "A first step towards in silico neuronal implmentation of early-vision map," in Proceedings of the Systems, Signal and Devices Conference (SSD 2007), Hammamet, Tunisia, 2007 (fulltext.pdf)
  • A. Daouzli, S. Saïghi, L. Buhry, Y. Bornat, and S. Renaud, "Weights Convergence and Spikes Correlation in an Adaptive Neural Network Omplemented on VLSI," in Proceedings of the International Conference on Bio-inspired Systems and Signal Processing (BIOSIGNALS), Funchal, Portugal, 2008, pp. 286-291 (fulltext.pdf)
  • J. Tomas, Y. Bornat, S. Saighi, T. Levi, and S. Renaud, "Design of a modular and mixed neuromimetic ASIC," in Proceedings of the 13th IEEE International Conference on Electronics, Circuits and Systems, (ICECS'06), Nice, France, 2006, pp. 946-949 (abstract)
  • S. Renaud, J. Tomas, Y. Bornat, A. Daouzli, and S. Saïghi, "Neuromimetic ICs with analog cores: an alternative for simulating spiking neural networks " in Proceedings of the IEEE 2007 InternationaI Symposium on Circuits And Systems (ISCAS'07), New-Orleans, USA, 2007, pp. 3355-3358 (fulltext.pdf)
  • T. Lévi, J. Tomas, N. Lewis, and P. Fouillat, "IP-based design reuse for analogue systems: a case study," in Proceedings of SPIE VLSI Circuits and Systems III, Maspalomas, Spain, 2007 (fulltext.pdf)


19 Feb 2010