Smart Networks and the Technophysics of Blockchain and Deep Learning

This article is a summary of our paper Swan M & dos Santos RP. Smart Network Field Theory: The Technophysics of Blockchain and Deep Learning, which was submitted to Concurrency and Computation: Practice and Experience. A preliminary version is available at arXiv.

Smart networks

Most progress to date, from mainframe to mobile, has focused on the transfer of basic information on simple networks. Now, however, in a second phase of network computing, a new paradigm is being inaugurated, smart networks, for the transfer of intelligent processing (see figure above). We foresee smart networks as intelligent autonomous networks, a new form of global computational infrastructure in which intelligence is built directly into the software, with an increasing degree of autonomous operation. Whereas previously the hardware pipes and the network architecture were key focal points in network analysis, now it is the intelligent software that operates the networks.

Blockchain distributed ledgers and deep learning systems are some of the most prominent examples of emerging smart network technologies (Table below).

Notice, however, that smart network technologies may also include different types of structural architecture and diverse operational objectives. Blockchains (transaction blocks cryptographically linked together) are one topology among alternative cryptographic structures that have been proposed. For example, “Blockless blocks” in the form of directed acyclic graphs (DAGs) have been proposed in merging projects including IOTA, Hashgraph, Byteball, and DAGCoin.[1]

Furthermore, several other kinds of smart network technologies (autonomous software operating over networks) already exist and have been operating for some time. These include unmanned aerial vehicles (drones), particularly those that have autonomous strike capability,[2] programmatic or high-frequency trading (HFT) which now comprises 55 percent of U.S. equities trading volume,[3] the real-time bidding (RTB) market for advertising, and smart energy grids with automated load-rebalancing. Other emerging and envisioned smart network technologies include smart city Internet of Things (IoT) sensor ecologies, automated supply chain logistics networks, cloudmind industrial robotics coordination networks (cloud-connected smart machines),[4] personal robotic assistant networks, and space-spaced logistics networks.

Smart Networks Types and Operational Focus

Deep Learning Chains

Smart network technologies may be used in convergence. Some species of smart network technologies, such as blockchain and deep learning, are in a special class in that they are both smart network technologies themselves and can serve as control technologies for other smart network systems. The functionality of blockchains and deep learning converges in the concept of deep learning chains. Deep learning chains are a smart network control technology with properties of both blockchain and deep learning: the secure automation, audit-log tracking, remunerability, and validated transaction execution of blockchains, and the object identification (IDtecha), pattern recognition, and optimization technology of deep learning.

Deep learning chains might be used to control other fleet-many internet-connected smart network technologies such as UAVs, autonomous driving fleets, medical nanorobots, and space-based asteroid mining rigs. For example, deep learning chains could be important in autonomous driving fleets for tracking what the vehicle does (blockchain), and for identifying objects in its driving field (deep learning). Deep learning chains could likewise apply to the body, as a smart network control technology for medical nanorobots, identifying pathogens (deep learning) and tracking and expunging them (blockchain). Likewise, in supply chain automated receiving, deep learning chains could provide the integrated functionality of object recognition and validated transfer.[5] An example of deep learning chains in operation is Provenance’s food supply chain traceability and attribute tracking system, which uses mobile apps, IoT sensors, blockchain, and machine learning. Sophisticated smart network control technologies such as deep learning chains require a strong theoretical model for their design and operation, and this is one motivation for the smart network field theory development proposed by this work.

Technophysics

On the other hand, Technophysics is the application of physics concepts and methods to the domain of technology, particularly for the study of smart network technologies, algorithms, and computation.6 The complex behaviour and fast-paced emergence of smart network technologies such as blockchain economic networks and deep learning pattern recognition systems suggest that rigorous methods for their study is needed, and that Physics would be an appropriate theoretical foundation.

Figure below helps to understand its definition, scope, and research topics by means of an analogy to Biophysics and Econophysics.

Technophysics: Definition, Scope, and Research Agenda

A key objective of Technophysics is deriving standardized methods for assessing complex system criticality and phase transition, as well as defining interim system structure between the levels of microscopic noise and macroscopic aggregates.

This discipline includes the complexity theoretic analysis of blockchain consensus algorithms to assess system criticality such as chaoticity and flash crashes.[6],[7] Other work theorizes smart networks,[8] articulating the smart network resource stack,[9] and aligning parallel formalisms in blockchain, deep learning, and black swan finance such as logistic regression, convexity, and programmable risk.[10]

A special form of smart network field theory has been proposed for the management of medical nanorobots in the biological setting.[11] In deep learning, technophysical approaches have been used to facilitate system convergence with spin glass models, for energy landscape minimization.[12] Deep learning has also been applied to the study of particle physics.[13] Complexity theory has been investigated for the control of another smart network system, drones or Unmanned Aerial Vehicles (UAVs), which exhibit undesirable emergent behaviour such as thrashing, resource starving, and phase change.[14]

Smart Network Field Theory

Large-scale networks are a feature of contemporary reality. They are complex systems comprised of thousands, millions, and billions of elements, and thus require a smart network field theory or other similar mechanism for the automated characterization, monitoring, and control of their activity, particularly for criticality detection and fleet-many item orchestration.

A theoretically-grounded model is necessary, and Swan proposes Smart Network Field Theory as a field theory based on statistical physics, effective field theories, and model systems. Other models might likewise be developed for the analysis and control of large-network systems. Blockchain and deep learning are chosen as the focus for developing a field theory since they are among the most sophisticated, robust, and conceptually novel.

Characterization is necessary to develop standard indicators and metrics to identify easily specific behaviours in smart network systems as they evolve and possibly grow in scalability and deployment. Both positive (emergent innovations) and negative (flash crash) behaviours should be assessed.

Monitoring, at both the individual element and overall system level, of current and evolving behaviour, pertains to smart network operations currently unfolding, and farther future needs. Although deep learning networks are currently isolated and restricted to certain computational infrastructures, it is not unimaginable in the farther future that deep thinkers (advanced deep learning systems) might be introduced to the internet. A Deep Thinkers Registry could be an obvious safeguard for tracking an entity’s activity, with possible annual review by a Computational Ethics Review Board for continued licensing. This is a far future example but demonstrates the intended extensibility of smart network field theories, and the uncertain future situations that they might help orchestrate.

Control, both for the coordination of fleet-many items, and for predictive risk management, is the third objective of a smart network field theory. The ability to coordinate fleet-many items in any kind of internet-connected smart network system, which could include autonomous vehicles, drones, blockchain p2p nodes, deep learning perceptrons, smart city IoT sensor landscapes, home-based social robots, medical nanorobots, and supply chain shipment-receiving, is also an obvious automation economy benefit of smart network technologies. The longer-term range of deployment of smart network technologies could extend to the very small, the cellular domains of the body, and the very large, such as terraforming in space.

The term field is meant here more analogically than literally (in the precise physical sense of an electromagnetic or gravitational field), and likewise other terms subsequently invoked here such as temperature and pressure, which also have precise meanings in the physical context. These terms are applied conceptually as to the purpose and function they are serving in physical systems, which is then extended to the context of smart network field theory development.

There are two primary meanings of field in the more conceptual smart network field theory sense. First and most generally, field is referring to the ability to control fleet-many items as one unit. This concept of field might be used to coordinate thousands and millions of constituent elements (such as blockchain p2p nodes or deep learning perceptrons). An example of an existing smart network field operation is optogenetics (in which neurons express a protein that makes their electrical activity controllable by a light switch as a field as opposed to individually).[15],[6]

The second meaning of field might refer to the situation in which each element in a system has its own measure and contribution to overall network activity (and can be used to calculate a Hamiltonian or other system composite measure). This concept of field (from effective field theory development in physics) suggests that every point in a landscape has a value which may be calculated.

A key requirement for a smart network field theory is that it can be used to manage across diverse system scales. Such a theory should be able to “identify macroscopic smoothness from microscopic noise” as complexity theory articulates.[16] Statistical mechanics and effective field are therefore selected as two methods for linking multiple dimensions within systems.

Statistical mechanics is an important generalized method, rooted in probability, for linking microscopic noise to macroscopic labels.[17] Thus, statistical mechanical field theories could be good models for the formalization of smart networks. The thermodynamics aspect of statistical mechanics is also an important and obvious formulation for the study of smart networks. Concepts such as work, heat, and energy have analogues between thermodynamical systems and smart networks. Macroscopic statistical quantities thus make sense in that there is a state of thermodynamical equilibrium in both blockchain systems and deep learning networks. In blockchain systems, equilibrium may be constituted different ways, including by a match between the replicability of information and the competition among peers writing to the ledger.[18] In deep learning networks, the system is likewise at equilibrium, finding an optimized solution after some number of forward and backward propagations.[19]

The other arm of smart network field theory development is effective field theories, a type of approximation, or effective theory, which explains an underlying domain in simpler terms, being used in basic physics, particularly quantum mechanics and materials science.[20] One of the most prominent examples of an effective field theory is the elliptical orbits of the planets, which are more easily calculated with Newtonian gravity than with general relativity.

An effective field theory is a formal process that can be used to identity analogues to macroscopic temperature and pressure terms corresponding to measures of microscopic behaviour in multi-level systems. Temperature is the consolidated measure that describes the movement of all of the particles in a room. Likewise, in a field theory, a temperature term is a consolidated measure that naturally occurs in any system in which there is a conserved quantity that comprises all of the activity of a microscopic level at one or more tiers down in the system. For example, in a biological neural network, the system-wide quantity of interest might be the spiking activation (the threshold at which neurons fire), and other data would be superfluous. temperature terms are macroscopic

The temperature term is chosen for being informative and might be employed as an overall system assessment parameter. As a practical example, if the pressure rises above some working range or the temperature drops below it, it is an indication of some possibly dangerous malfunction of the system that must be overseen. An informative measure for blockchain systems would be one that warned, for example, that a 51% or DDoS attack was forming, or that there are so few online nodes that the security of the network is compromised.

The Smart Network Field Theory is developed qualitatively, not algorithmically, with close association to two existing field theory systems, Cowan’s Statistical Neural Field Theory for describing brain activity,[21] and Wolynes’s Spin Glass Model,[22] extended with LeCun’s theoretical work in deep learning systems,[23] to specify how protein-folding and deep learning systems converge on a solution.

It is crucial to have a mathematically- and theoretically-based smart network field theory for understanding smart networks for predictive risk management in being able to detect and possibly avert system criticality such as potential phase transitions in smart networks, in both expected and emergent situations such as financial contagion, network security, and electromagnetic pulses (either malicious or unintentional).

Practical applications of smart network field theories include the network nodes, states, actions, and phase transitions of the system.

Nodes (Particles)

In smart networks, the particles are the constituent network elements. In blockchain networks, the core unit is the p2p (peer to peer) node. This is the flat-hierarchy distributed p2p nodes that offer fee-based services for activities such as transaction validation and recording (mining). There are also other units in blockchain networks such as wallets, transactions, digital assets, and digital asset inventories. In deep learning networks, the constituent units are perceptron nodes, which are modular units, organized in a graph-theoretic structure having nodes and edges. Input values enter the processing node from the input edges, are processed by logic in the node, and exit as an output value on the output edge, which is used as the input value for a downstream node. Thousands and millions of nodes are organized Lego-like into cascading layers of processing (five to seven layers of thousands or millions of nodes on average).

State Transitions

An important goal of a field-theoretic formulation is to be able to identify system phase transitions, both positive and negative. The idea is to capture the system-wide calculation of the probability that certain nodes are in certain states, and thus provide resource indicators or economic indicators as a class of risk-assessment metrics. In a practical example, a smart network field theory could be used to understand and evaluate systemic risk and the possibility of flash crashes. This is particularly important now that trading is a multi-species activity, conducted half by human agents and half by programmatic trading in worldwide equities markets.11

Computational Complexity

One implication of smart network field theory is that it is a complexity method that can be applied to the consideration of computational complexity, and possibly offer alternative ways of casting the P/NP problem schema. Advance in rethinking computational complexity has been obtained from the consideration of quantum gravity[24] and black holes[25] as alternative kinds of information domains, and the advent of smart networks might sponsor the same. The native time and space domains of smart networks are distinct: temporally with block-time and learning network solve-time; and spatially with arbitrarily-expandable distributed p2p nodes, hidden processing layers, and self-optimizing architectures. Network computing may not have the same constraints of the traditional methods of evaluating P/NP-hard problems limited to running on von Neumann architectures and parallelized supercomputing architectures. Smart networks are, therefore, no longer subject to the traditional time and space constraints which structure the P/NP schema, which suggests recasting the problem schema based on other parameters may lead to advance in computational concurrency

Conclusion

The biggest potential practical benefit of smart network field theory is that it might be used to understand and manage smart network systems as they continue to evolve. Another potential benefit is using smart network field theories as the basis for designing, testing, and deploying new smart network systems, and for provisioning additional network layers and resources within existing smart network systems. Smart network field theories could lead to the well-formed engineering and instantiation of complex systems. A farther practical benefit of smart network field theories is that they might be used to identify novelty such as system criticality and phase transition, for risk management (for example, understanding and avoiding flash crashes), and for beneficial emergence (for example, articulating new systems-level resources such as algorithmic trust (blockchain) or time-series forecasting capability (deep learning)). Another practical benefit of this work is the informational resource created by the detailed summary of the model systems (Cowan’s Statistical Neural Field Theory and Wolynes’s Spin Glass Model) highlighted and theoretically derived in the extended version available at arXiv.

One of the farther theoretical consequences of this work is articulating the notion of smart network field theories as a means of assessing system criticality provides a foundation for the further study of this topic. The Technophysics concern of system criticality could become more of a formal area of study, as complexity theory has, with theoretical constructs and mathematical formalisms. System criticality and phase transition is of interest, but precursors for its avoidance are also important, especially from a complex systems management perspective. Smart network field theories might be helpful in eliciting interim system structure between the levels of microscopic noise and macroscopic labels.

References

[1] Lunn B. Will the Blockchain Economy be powered by something totally different like IOTA Tangle or Hashgraph? Daily Fintech. 2017.

[2] Vásárhelyi G, Virágh C, Somorjai1 G, et al. Optimized flocking of autonomous drones in confined environments. Science Robotics. 2018;3(20):eaat3536.

[3] Miller RS, Shorter G. High Frequency Trading: Overview of Recent Developments. U.S. Congressional Research Service. 2016(7-5700):R44443.

[4] Keenan T. CloudMinds, the World’s First Cloud Robot Operator, Launches Mobile-Intranet Cloud Services, Enabling Secure Cloud Robotic Deployments, and Data A1 Handset, World’s First Mobile Phone Robotic Control Unit. Business Wire. 2017.

[5] Swan M. Blockchain for Business: Next-Generation Enterprise Artificial Intelligence Systems. In: Raj P, Deka GC, eds. Advances in Computers. Vol. 111. Blockchain Technology: Platforms, Tools and Use Cases. London: Elsevier. 2018.

[6] dos Santos RP. On the Philosophy of Bitcoin/Blockchain Technology: Is it a Chaotic, Complex System? Metaphilosophy. 2017;48(5):620–633.

[7] dos Santos RP. Consensus Algorithms: A Matter of Complexity? In: Swan M, Potts J, Takagi S, Witte F, Tasca P, eds. Blockchain Economics. London: World Scientific. Forthcoming 2019.

[8] Swan M. Blockchain: Blueprint for a New Economy. Sebastopol CA: O’Reilly Media. 2015.

[9] Swan M. Decentralized Crypto-Finance: Blockchains, Automatic Markets, and Algorithmic Trust. Institute for Ethics and Emerging Technologies. 2016.

[10] Swan M. Blockchain Theory of Programmable Risk: Black Swan Smart Contracts. In: Swan M, Potts J, Takagi S, Witte F, Tasca P, eds. Blockchain Economics. London: World Scientific. Forthcoming 2019.

[11] Swan M. Technophysics, Smart Health Networks, and the Bio-cryptoeconomy: Quantized Fungible Global Health Care Equivalency Units for Health and Well-being. In Boehm, F. Ed., Nanotechnology, Nanomedicine, and AI: Toward the Dream of Global Health Care Equivalency. Boca Raton FL: CRC Press. Forthcoming 2019.

[12] Choromanska A, Henaff M, Mathieu M, Arous GB, LeCun Y. The loss surfaces of multilayer networks. Artificial Intelligence and Statistics. 2015;192-204.

[13] Radovic A, Williams M, Rousseau D, et al. Machine learning at the energy and intensity frontiers of particle physics. Nature. 2018;560:41-48.

[14] Singh S, Lu S, Kokar MM, Kogut PA. Detection and classification of emergent behaviors using multi-agent simulation framework (WIP). Spring 2017 Simulation Multi-Conference. Society for Modeling & Simulation (SCS). 2017.

[15] Boyden ES. Optogenetics and the future of neuroscience (Overview, Optogenetics 10th Anniversary Issue). Nature Neuroscience. 2015(18):1200–1201.

[16] Mitchell M. Complexity: A Guided Tour. Oxford UK: Oxford University Press. 2009.

[17] Mayants L. The enigma of probability and physics. New York: Springer. 1984;174.

[18] Abadi J, Brunnermeier M. Blockchain Economics. 2018. https://scholar.princeton.edu/sites/default/files/markus/files/blockchain_paper_v2c.pdf.

[19] Alemi AA, Fischer I. TherML: Thermodynamics of Machine Learning. arXiv:1807.04162v2 [cs.LG]. 2018.

[20] Giorgi GA, Guerraggio A, Thierfelder J. Mathematics of Optimization: Smooth and Nonsmooth Case. London: Elsevier. 2004.

[21] Cowan J. A Personal Account of the Development of the Field Theory of Large-Scale Brain Activity from 1945 Onward. In: Coombes S, Graben P, Potthast R, WrightJ, eds. Neural Fields: Theory and Applications. Heidelberg: Springer. 2014.

[22] Bryngelson JD, Onuchic JN, Socci ND, Wolynes PG. Funnels, pathways, and the energy landscape of protein folding: a synthesis. Proteins: Structure, Function, and Bioinformatics. 1995;21(3):167-195.

[23] LeCun Y, Bengio Y, Hinton G. Deep learning. Nature. 2015;521:436-444.

[24] Susskind L. Black Holes and Complexity Classes. arXiv:1802.02175 [hep-th]. 2018.

[25] Aaronson S. Why Philosophers Should Care About Computational Complexity. Electronic Colloquium on Computational Complexity. 2011.

---