The processing of information following the classical von Neumann digital computing paradigms is known to be less efficient compared to the biological counterparts, when dealing with ill-posed problems and noisy data. Though current computing technologies have reached speed and computational power figures that allows them to simulate parts of animal brains and behaviour, the energy required by these systems grows exponentially with the increasing hierarchy of animal intelligence (Jo et al., 2010). The reason is that the biological brain is configured differently and the keys are the extremely high (~1015 synapses) connectivity between neurons in a network which offers highly parallel processing power (Jo et al., 2010) as well as the fact that neurons are plastic and adaptive (i.e. memory dependent) signal processing and computing units. Yet, brain’s most striking feature is that it is structured as an evolving system were synapses undergo ‘birth’ and ‘death’ as well as strengthening and weakening, reconfiguring neuronal connectivity in a self-organizing manner and allowing the networked population of neuronal processors to adapt motor and behavioural responses to the ever changing environmental inputs. Thus, by rearranging both the structural and functional topology (Chiang and Yang, 2004), brain’s neuronal circuits demonstrate unique evolvability, scalability and adaptability properties that are unmatched by current computing devices. For years, the implementation of artificial neural networks has been in the form of software run on conventional “von Neumann” computers, and such simulations appear incomparable to the biological systems in terms of speed and energy efficiency. The challenge has been to propose a “physical neural network” where elements overcome this deficiency by merging data storage and processing into single electronic devices and where topology can be reconfigured in a self-reorganizing manner. Such artificial neural network could provide the complexity, connectivity, and massive parallel information processing and thus mimic the performance of biological systems including their evolvability, self-organization, adaptability and robustness.
Bio-inspired “neuromorphic” computation schemes rely, on one hand, on the deep knowledge of neuron functionality and on the other hand may take advantage of modern emerging low power nanoelectronic devices particularly suitable to emulate neuron/synapsis functional properties. Among the emerging technologies which may enable functional scaling of logic and memory circuits well beyond the limits of CMOS devices, a particularly promising concept is the memristor (memory – resistor) – a nanoscale thin-film device that maintains a functional relationship between the time integrals of current and voltage (Chua, 1971; Chua, 2011; Prodromakis et al., 2012). The non-linear dynamics as well as the plasticity of the newly discovered memristor are shown to support Spike-based- and Spike-Timing-Dependent-Plasticity (STDP), making this extremely compact device an excellent candidate for realizing large-scale self-adaptive circuits; a step towards “autonomous cognitive systems”.
In our project, the intrinsic properties of real neurons and synapses as well as their organization in forming neural circuits will be exploited for optimizing CMOS-based neurons, memristive grids and the integration of the two into realtime biophysically realistic neuromorphic systems. The artificial neuromorphic platforms will be interfaced to real neurons from the rat brain through electrical nano-/micro-sensors and stimulators forming an evolving and self-adapting biohybrid system. The biohybrid technological platform will exploit at the elementary level the two basic features responsible for brain evolvability and adaptability: synaptic plasticity and reconfigurability of neuronal wiring. During the project, memristive plasticity will be optimized and merged with the natural counterpart in bio-hybrid natural-artificial (wetware-hardware) neuronal microcircuits. Evolving and self-organizing connectivity among artificial neurons in the network will be implemented taking the cue from neuronal connections as they occur in the real network.
Throughout our work we will learn from real neurons how to embed evolving connectivity and plasticity figures in CMOS/memristive architectures for brain inspired computing. On the other hand, the engineering of the neuromorphic system will lead to a better understanding of the biophysical basis of information processing in real neuronal circuits. Overall the new system will provide a first proof-of-concept that neurons can mutually interact with nanoelecronic memristive devices sharing similar memory and plasticity rules, co-evolving and adapting to external world inputs.