PhD proposal: Bio-inspired self-organizing neural architectures
Bernard Girau (Bernard.Girau@loria.fr)
Team : BISCUIT, LORIA
neural networks, self-organization, neural fields, reconfigurable circuits
In the context of a collaboration ranging from computational neurosciences to the hardware design of reconfigurable multicore circuits, the main goal of this PhD project is to study the possible contributions of bio-inspired principles of structural neural plasticity in the field of computing architecture design, focusing on the definition of neuro-cellular models able to both represent external data and adapt their own underlying structure to hardware constraints.
As the quest for performance in computing systems confronts resource constraints, major breakthroughs in computing efficiency are expected to benefit from unconventional approaches and new models of computation such as brain-inspired computing. The brain is a massively parallel architecture with dense interconnections between computing units. Its structural organization is already a source of inspiration of several neuromorphic integrated circuits, but several dynamic properties of the brain computation have not really been explored from the point of view of a computational architecture. We are interested in the design of adaptive and dynamic computing architectures, taking advantage of brain-inspired self-organization principles. Self-organizing neural models have already been studied, especially for vector quantization tasks. However, existing self-organizing models are limited in the context of a self-organizing computing architecture project, because they rely on a predefined topology and they do not take into account communication costs nor afferent connectivity. Therefore, 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. Structural plasticity has been hardly considered in most self-organizing models because on one hand, such explicit dynamic lateral connectivity is computationally expensive and on the other hand, the exact rules governing the growing or the pruning of connections remain unclear. In collaboration with researchers involved in the definition of computational models of the biological brain, the first part of this thesis project will consist in:
– extending the existing capabilities of neural self-organizing models so that bio-inspired processes of structural plasticity are added to the more usual mechanisms of synaptic plasticity
– interpreting the resulting classification in terms of communication requirements among the computing units of the self-organizing architecture
Moreover, we want our models to be fault tolerant and scalable. 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. We also want our models to be hardware-compliant, thanks to a neuro-cellular computing substrate. The goal of the second part of this thesis project is thus to adapt the multilevel self-organization models that will derive from our bio-inspired approach to the constraints of a cellular computing substrate using neural spikes. The main challenge is there to define purely local rules of structural plasticity within these models, for example by using Hebbian/STDP-like learning rules for synaptic as well as structural plasticity in our neuro-cellular models.
Profile and skills required
The candidate should have the equivalent of a Master in Computer Science, preferably in a specialty related to artificial intelligence and/or distributed numerical computation and/or bio-inspired computation. Adequate knowledge of digital hardware design will be assessed, as well as experience in software design. The candidate must fluently speak English and/or French.
 M. Sipper, “The emergence of cellular computing,” Computer, vol. 32, no. 7, pp. 18–26, 1999.
 T. Kohonen, “Self-organized formation of topologically correct feature maps,” Biological Cybernetics, vol. 43, 1982.
 B. Chappet De Vangel and B. Girau, “Stochastic and asynchronous spiking dynamic neural felds,” in International Joint
Conference on Neural Networks, IJCNN, 2015.
 A. Upegui, Y. Thoma, A. Perez-Uribe, and E. Sanchez, “Dynamic routing on the ubichip: Toward synaptogenetic neural
networks,” in Adaptive Hardware and Systems. IEEE, 2008, pp. 228–235.
 L. Rodriguez, B. Miramond, and B. Granado, “Toward a sparse self-organizing map for neuromorphic architectures,” ACM
Journal on Emerging Technologies in Computing Systems, vol. 11, no. 4, p. 33, Apr. 2015.
 T. Rumbell, S. Denham, and T. Wennekers, “A Spiking Self-Organising Map Combining STDP, Oscillations and Continuous
Learning,” IEEE TRANSACTIONS ON NEURAL NETWORKS AND LEARNING SYSTEMS, vol. 5, 2014.