Networks of "silicon" neuronsIn collaboration with Sylvie Renaud and Sylvain Saïghi (Université de Bordeaux), we have been conceiving systems for simulating networks of neurons using analog integrated circuits (ASICs). This project was supported by a European grant (SenseMaker) and continued under two European Integrated Projects (FACETS and BrainScales). Together with Renaud's laboratory for microelectronics, we have designed a system to simulate networks of analog HodgkinHuxley type neurons where the connectivity is handled by a computer. The system was entirely realized in Bordeaux by Renaud's group, specialized in microelectronics. This solution offers to solve in hardware what is slow to solve using a computer, namely the membrane equations, HodgkinHuxley dynamics of conductances, synaptic inputs, etc. In contrast, the computer handles the connectivity between neurons, which is typically hard to integrate in hardware. The circuits that we have designed were realized by the laboratory of Renaud, as well as the electronics around the system and its (realtime) interface with the computer. The role of our group has been to initiate this project, participate to the design, and give the neuron models that were integrated on the circuits, as well as build different configurations and run the models. We have chosen to include two main classes of cortical neurons, "regularspiking" (RS, excitatory) and "fastspiking" (FS, inhibitory), which were described by HodgkinHuxley models. Synaptic inputs are mediated by conductancebased excitatory (AMPA) and inhibitory (GABA_{A}) synapses. We used a kinetic model of synapse [1] that was directly implemented in the circuit. The connectivity is entirely programmable via the computer interface, and all synaptic weights dynamics are also handled by the computer, so this system can simulate any type of network of RS and FS cells, with any type of plasticity algorithm. This first prototype consisted of a network of 8 analog HodgkinHuxley neurons [2]. This prototype was tested in in various ways [2,3]. First, at the singlecell level, we compared the ASIC neurons with computer simulations, using different protocols, such as the firing frequency with constant current injection (fi curve), and irregular firing following random synaptic bombardment. In both cases, the ASIC neurons matched numeric simulations. The system was then tested using small networks of excitatoryinhibitory neurons, with synapses following spiketiming dependent plasticity (STDP). With this severe test, the matching between ASIC and numeric models was excellent (see details in [3]). More recently, we collaborated with the groups of Sylvain Saïghi and Sylvie Renaud to a second generation system aimed to include of the order of 100 ASIC neurons [4]. This system should be able to simulate a larger repertoire of neuronal classes, including RS and FS cells and bursting neurons. We have independently characterized HodgkinHuxley type models of various neuron types [5], and our goal is to include these cell classes into networks on hardware. A first step towards this goal was obtained recently [6]. In collaboration with the group of Karlheinz Meier (University of Heidelberg), we have participated to a collective work (within the projects FACETS, BrainScales and the Human Brain Project), based on a different type of VLSI hardware aimed at simulating largescale networks of integrateandfire neurons. We implemented networks of adaptive exponential integrate and fire neurons and succeeded in simulating in the hardware the different AI states observed in model networks. This work was part of the final "demonstrators" of the FACETS and BrainScales projects and is now continued in the Human Brain Project [7, 8]. [1] Destexhe, A., Mainen, Z.F. and Sejnowski, T.J. An efficient method for computing synaptic conductances based on a kinetic model of receptor binding. Neural Computation 6: 1418, 1994 (see abstract) [2] Zou, Q., Bornat, Y., Tomas, J., Renaud, S. and Destexhe, A. Realtime simulations of networks of HodgkinHuxley neurons using analog circuits. Neurocomputing 69: 11371140, 2006 (see abstract) [3] Zou, Q., Bornat, Y., Saïghi, S., Tomas, J., Renaud, S. and Destexhe, A. Analogdigital simulations of full conductancebased networks of spiking neurons with spike timing dependent plasticity. Network 17: 211233, 2006 (see abstract) [4] Renaud, S., Tomas, J., Lewis, N., Bornat, Y., Daouzli, A., Rudolph, M., Destexhe, A. and Saïghi, S. PAX: A mixed hardware/software simulation platform for spiking neural networks. Neural Networks 23: 905916, 2010 (see abstract) [5] Pospischil, M., ToledoRodriguez, M., Monier, C., Piwkowska, Z., Bal, T., Fregnac, Y., Markram, H. and Destexhe, A. Minimal HodgkinHuxley type models for different classes of cortical and thalamic neurons. Biol. Cybernetics 99: 427441, 2008 (see abstract) [6] Grassia, F., Buhry, L., Levi, T., Tomas, J., Destexhe, A., and Saighi, S. Tunable neuromimetic integrated system for emulating cortical neuron models. Frontiers in Neuroscience 5: 134, 2011 (see abstract) [7] Bruderle D, Petrovici MA, Vogginger B, Matthias Ehrlich M, Pfeil T, Millner S, Grubl A, Wendt K, Muller E, Schwartz MO, Husmann de Oliveira D, Jeltsch S, Fieres J, Schilling M, Muller P, Breitwieser O, Petkov V, Muller L, Davison AP, Krishnamurthy P, Kremkow J, Lundqvist M, Muller E, Partzsch J, Scholze S, Zuhl L, Mayr C, Destexhe A, Diesmann M, Potjans TC, Lansner A, Schuffny R, Schemmel J and Meier K. A comprehensive workflow for generalpurpose neural modeling with highly configurable neuromorphic hardware systems. Biol. Cybernetics 104: 263296, 2011 (see abstract) [8] Petrovici, M.A., Vogginger, B., Muller, P., Breitwieser, O., Lundqvist, M., Muller, L., Ehrlich, M., Destexhe, A., Lansner, A., Schuffny, R., Schemmel, J. and Meier, K. Characterization and compensation of networklevel anomalies in mixedsignal neuromorphic modeling platforms. PLoS One 9: e108590, 2014 (see abstract)
Department of Integrative and Computational Neuroscience (ICN),
