SOCRATES: Self-Organising Computational substRATES

The growth of data will influence all information processing system, from the simplest sensor node to the most complex supercomputer. Contemporary computer systems struggle with energy efficiency and basic architectural to meet the challenge of today's vast data growth. SOCRATES is a long-term time horizon project seeking radical breakthroughs toward efficient and powerful data analysis available everywhere. SOCRATES will take inspiration from biology and neuroscience to exploit novel substrates that support self-organization through local interactions to create a theoretical and experimental foundation for a new computing paradigm. In the SOCRATES project we exploit the self-organizing properties of living neural cultures as a method for creating new information processing hardware in nanomagnetc substrates. Living neural cultures of SOCRATES are grown to collect data on how such network organize. Data from current neural cultures have been collected and are analysed to detect network properties related to self-organization, i.e. network structure and behaviour. Models of network behaviour is used to investigate the computational abilities of these biological systems toward incorporation in our targeted nanomagnetic computing hardware. The SOCRATES team has found models and produced simulation tools that can model our complex systems at a realistic level and speed that enable system of a scale necessary for computing. Nanomagnet ensembles has been fabricated at NTNU NanoLab. Nanosystems consisting of up to approx. 100 000 nanomagnets with several geometries (architectural layouts) have been produced. To be able to measure behaviour and show that our simulation results can be replicated in physical systems. We have used international laboratories including synchrotron facilities (X-PEEM) and local laboratory facilities at NTNU (MFM and MOKE). For our living neural networks human engineered neuronal subtypes and neurons derived from rodent brain has been used to studied emergent behaviour, including the possibility to influence structural and/or functional dynamics. The SOCRATES project ended in 2022. At the end of the project the goal of harnessing emergent and self-organizing properties for computation is demonstrated within the reservoir computing framework. An artificial neural network, Reservoir Computation, that including the possibility of feedback, and there by also memory, is as such the computational model of the SOCRATES project used to show computation in nanomagnet ensembles. To be able to take the complex behaviour of neural systems, or any such system operating at continuous time and continuous space, to an abstraction suitable for computing, SOCRATES have made a discrete abstraction. A simulation framework to investigate discrete dynamic models of self-organizing systems called EvoDynamics, an open-source simulator. Such a simulator enables SOCRATES to model physical systems at a level that suits the chosen reservoir computational model. EvoDynamic is designed to target any sparsely connected discrete network to explore network structure, local communication and learning rules. The abstraction of discrete dynamical systems is reflected in the nanomagnet computational substrate. SOCRATES uses Artificial Spin Ice (ASI), a nanomagnet system where each nano magnet is engineered to be a dipole, i.e. a magnet’s magnetization can only point in one of two directions, and therefore also can be seen as a binary element. To be able to explore different ASI geometries and investigate protocols for stimulation and tuning a realistic model for such a dipole approximation was necessary to be able to simulate ASI systems within the reservoir computational framework. The open source flatspin simulator is the SOCRATES project’s nanomagnet simulator and modelling tool. flatspin can model the discrete dynamics of large scale nanomagnet ensembles at a level that are reproducible in physical experiments. As such, the SOCRATES project has been able to show that ASI based nanomagnet substrates can be tuned to do useful computation. In SOCTRATES the discrete dynamic abstraction for complex dynamic systems has been used to show that ASI is a very promising substrate for computation. Experimental results, in simulation, has shown that ASIs can be tuned to perform AI related tasks, not only at a single task, but at a general level demonstrated by solving benchmark tests that include general properties that are necessary for all classification tasks. The benchmark tests also show that the property of fading memory, a necessary property for temporal tasks, is present in ASI systems. Further, phenomena, protocols, and tuning methods for controlling behaviour and capturing realistic dynamics, in ASIs at long time scales has been found in simulation and verified in physical experiments

Primary: -Open-source models and simulator framework: Two open-source for i) simulation of physical reservoir as discrete dynamic system, called "EvoDynamics". ii) a dipole model for ASI including a simulation framwork capable of simulating large scale discrete dynamics at a level of accuracy that can be verified experimentally, called "flatspin". The two models and simulator framework are both based on models of interacting elements with a possibility to perturb. -Open access data from in vitro MicroElectrode Array experiments. -Protocols and methods to tune nanomagnet ensembles to dynamic regimes where computational properties can be exploited.. -Prove of principle of computation in nanomagnet systems using the Reservoir Computing paradigme.. -ASI as a tuneable reservoir for the reservoir computing framework -Showing computational properties in ASIs that are significant for AI type of problems, i.e. classification and pattern recognition and memory. Results are general, The well known benchmark of generalization and kernel quality together with a measure of computational quality was used to show that ASI based nanosystems indeed have required properties to do useful computation. Secondary: -Of the four PhD candidates one have delivered, the thesis is approved and the defence is set to March 13th . One is currently on parental leave, writing up stage, expect to deliver in 2023. One has been granted extra time to be able to have a sty abroad, also in write-up stage, expected to deliver 2023. The last is also in the write-up stage and expected to deliver in 2023. The project manage good in the Coronavirus period, but the PhDs publishing was is delayed. In the project period 25 academic papers has been published, there are 3 in process near publication in 2023 and results that need more processing and confirmation experiments for later publications. In the project period the SOCRATES team has strengthen the collaboration with University of York and University (UK) of Ghent (BE). There as been established a close and active collaboration with the University of Sheffield (UK), the Paul Scherrer Institute at ETH-Zurich and IBM research Zurich. The SOCRATES project can be said to be an early initiative to the newfound interest for morphogenetic engineering in physical substrates and the revisited focus on unconventional computing. Internationally the European union has supported several projects, e.g. HYBRAIN where SOCRATES team member is an external monitors. A SOCRATES team member act as party of the UK funded MARCH project, coordinated by the University of York. The team has given key notes at international conferences and workshops. Parts of the SOCRATES team successfully applied for a FET-Open project, SpinENGINE. The project is coordinated by NTNU. Partners are University of Sheffield, the University of Ghent, the Paul Scherrer Institute at ETH-Zurich and IBM research Zurich.

SOCRATES is a long-term time horizon project seeking radical breakthroughs toward efficient and powerful data analysis available everywhere, from the simplest sensor node to the most complex supercomputer. SOCRATES will exploit novel substrates that support self-organization through local interactions to create a theoretical and experimental foundation for a new computing paradigm. Such a complex systems approach to analytics opens for a radical breakthrough in the field of computing, alleviating main problems of contemporary computer systems relating to energy efficiency, scalability, and self-learning. The data analytic challenge is importunate in today's increasingly data-rich society. Where a staggering 2.5 exabytes are created every day and emerging technologies like the Internet of things (IoT) will substantially increase the data growth rate, and further increase the demand for efficient analysis. To achieve efficient analysis everywhere, fundamentally new hardware approaches that are efficient, scalable, and may be adapted to the needs of diverse and complex data analysis tasks are required. An ideal system for realization of efficient hardware should be capable of vast parallel processing of data with inherent parallel learning capability. In SOCRATES, a fundamentally new hardware approach based on principles from morphogenetic engineering (i.e., dynamic development of complex heterogeneous structures) will be developed to meet these requirements. SOCRATES will leverage substrates with self-organizing and emergent behaviour to create systems with the property of inherently changing state transition functions and the set of state variables over time (caused by bio-inspired morphological processes). We aim at creating a theoretical and experimental foundation of morphogenetic systems based on self-organizing and emergent behaviour in biological neural nets and ensembles of nanomagnets that have all the desired properties of an ideal system for data analysis.