Why Machine Learning and the ‘New AI’ won’t be Replacing your Friendly Post – Keynesian Macroeconomist Anytime Soon


The paper provides a brief history of recent developments in machine learning and the “New AI”.  This sets the scene for a review of debates over machine learning and scientific practice, which brings to the forefront the hubris of those appealing to a naïve form of materialism in this specific domain at the intersection between philosophy and sociology of science. The paper then explores the “unreasonable effectiveness” of machine learning to shine a spot-light on the limitations of contemporary techniques. The resulting insights are subsequently applied to the particular question of whether current machine learning platforms could capture key elements responsible for the complexity of real-world macroeconomic phenomena as these have been understood by Post Keynesian economists. After concluding in the negative, the paper goes on to examine whether efforts to extend deep learning through differential programming could overcome some of the previously discussed limitations and stumbling blocks.

Keywords: machine learning, the “New AI”, macroeconomic modelling, fixed-point theorems, backpropagation, the capital debates, uncertainty, financial instability, differential programming


An avalanche of recent publications (Zuboff, 2019; Gershenfeld, Gershenfeld & Gershenfeld, 2017; Carr, 2010; Lovelock, 2019; and Tegmark, 2017) reflect the emotional range of our current obsessions about the Digital Economy, which are concerned, respectively, with: its inherent capacity for surveillance, domination, and control; its opportunities for extending the powers of digital fabrication systems to all members of the community; its retarding effects on deep concept formation and long-term memory; the prospect of being watched over by “machines of loving grace” that control our energy grids, transport and weapon systems; and, the limitless prospects for the evolution of AI, through procedures of “recursive self-improvement”. In my own contribution to the analysis of the digital economy (Juniper, 2018), I discuss machine learning and AI from a philosophical perspective that is informed by Marx, Schelling, Peirce and Steigler, arguing for the development of new semantic technologies based on diagrammatic reasoning, that could provide users with more insight and control over applications.[1]

AI and Machine Learning practitioners have also embraced the new technology of Deep Learning Convolution Neural Networks (DLCNNs), Recursive Neural Networks, and Reservoir Neural Networks with a mixture of both hubris and concern[2]. In an influential 2008 article in Wired magazine, Chris Anderson claimed that these new techniques no longer required a resort to scientific theories, hypotheses, or processes of causal inference because the data effectively “speak for themselves”. In his response to Anderson’s claims, Mazzochi (2015) has observed that although the new approaches to machine learning have certainly increased our capacity to find patterns (which are often non-linear in nature), correlations are not all there is to know. Mazzochi insists that they cannot tell us precisely why something is happening, although they may alert us to the fact that something may be happening. Likewise, Kitchin (2014) complains that the data never “speak for themselves”, as they shaped by the platform, data ontology, chosen algorithms and so forth. Moreover, not only do scientists have to explain the “what”, they also have to explain the “why”. For Lin (2015) the whole debate reflects a confusion between the specific goal of (i) better science; and that of, (ii) better engineering (understood in computational terms). While the first goal may be helpful, it is certainly not necessary for the second, which he argues has certainly been furthered by the emerging deep-learning techniques[3].

In what follows, I want to briefly evaluate these new approaches to machine learning, from the perspective of a Post Keynesian economist, in terms of how they could specifically contribute to a deeper understanding of macroeconomic analysis. To this end, I shall investigate thoughtful explanations for the “unreasonable effectiveness” of deep-learning techniques, which will therefore focus on the modelling, estimation, and (decentralised) control of system (-of systems) rather than image classification or natural language processing.

The “Unreasonable effectiveness” of the New AI

Machine learning is but one aspect of Artificial Intelligence. In the 1980s, DARPA temporarily withdrew funding for US research in this field because it wasn’t delivering on what it had promised. Rodney Brooks has explained that this stumbling block was overcome by the development of the New AI, which coincided with the development of Deep Learning techniques characterised by very large neural networks featuring multiple hidden layers and weight sharing. In Brooks’ case, the reasoning behind his own contributions to the New AI were based on the straightforward idea that previous efforts had foundered on the attempt to combine perception, action, and logical inference “subsystems” into one integrated system. Accordingly, logical “inference engines” were removed from the whole process so that system developers and software engineers could just focus on more straightforward modules for perception and action. Intelligence would then arise spontaneously at the intersection between perception and action in a decentralized, but effective manner.

One example of this would be the ability of social media to classify and label images. Donald Trump could then, perhaps, be informed about those images having the greatest influence over his constituency, without worrying about the truth-content that may be possessed by any of the individual images (see Bengio et al., 2014, for a technical overview of this machine learning capability). Another example of relevance to the research of Brooks, would be an autonomous rover navigating its way along a Martian dust plain, that is confronted by a large rock in its path. Actuators and motors could then move the rover away from the obstacle so that it could once again advance unimpeded along its chosen trajectory—this would be a clear instance of decentralized intelligence!

In their efforts to explain the effectiveness of machine learning in a natural science context, Lin, Tegmark, and Rodnick (2017), consider the capacity of deep learning techniques in reproducing Truncated Taylor series for Hamiltonians.  As Poggio et al., (2017) demonstrate, this can be accomplished because a multi-layered neural network can be formally interpreted as a machine representing a function of functions of functions… :


At the end of the chain we arrive at simple, localized functions, with more general and global functions situated at higher levels in the hierarchy. Lin, Tegmark, and Rodnick (2017) observe that this formalism would suffice for the representation of a range of simple polynomials that are to be found in the mathematical physics literature (of degree 2-4 for the Navier-Stokes equations or Maxwell’s equations). They explain why such simple polynomials characterise a range of empirically observable phenomena in the physical sciences, in terms of three dominant features, namely: sparseness, symmetry, and low-order[4]. Poggio et al., (2017) examine this polynomial approximating ability of DLCNNs, also noting that sparse polynomials are easier to learn than generic ones owing to the parsimonious number of terms, trainable parameters, and the associated VC dimension of the equations (which are all exponential in the number of variables). The same thing applies to highly variable Boolean functions (in the sense of having high frequencies in their Fourier spectrum). Lin, Tegmark, and Rodnick (2017) go on to consider noise from a cosmological perspective, noting that background radiation, operating as a potential source of perturbations to an observed system, can be described as a relatively well-behaved Markov process.

In both of these cases, we can discern nothing that is strictly comparable with the dynamics Post Keynesian theory, once we have abandoned the Ramsey-Keynes (i.e. neoclassical) growth model as the driver of long -run behaviour in a macroeconomy. From a Post Keynesian perspective, the macroeconomy can only ever be provisionally described by a system of differential equations characterised by well-behaved asymptotic properties of convergence to a unique and stable equilibrium.

The Macroeconomy from a Post Keynesian Perspective:

In The General Theory, Keynes (1936) argued that short-run equilibrium could be described by the “Point of Effective Demand”, which occurs in remuneration-employment space, at the point of intersection between aggregate expenditure ( in the form of expected proceeds associated with a certain level of employment) and aggregate supply (in the form of actual proceeds elicited by certain level of employment). At this point of intersection, the expectation of proceeds formed by firms in aggregate is fulfilled, so that there is no incentive for firms to change their existing offers of employment. However, this can occur at a variety of different levels of employment (and thus unemployment).

For Keynes, short-run equilibrium is conceived in terms of a simple metaphor of a glass rolling on a table rather than that of a ball rolling along in a smooth bowl with a clearly defined minimum. When it comes to the determination of adjustments to some long-run full-employment equilibrium, Keynes was no less skeptical. Against the “Treasury-line” of Arthur Pigou, Keynes argued that there were no “automatic stabilizers” that could come into operation. Pigou claimed that with rising unemployment wages would begin to fall, and prices along with them. This would make consumers and firms wealthier in real terms, occasioning a rise in aggregate levels of spending. Instead, Keynes insisted that two other negative influences would come into play, detracting from growth. First, he introduced Irving Fisher’s notion of debt-deflation. According to Fisher’s theory, falling prices would transfer income from high-spending borrowers to low-spending lenders, because each agent was locked in to nominal rather than real or indexed contracts. Second, the increasing uncertainty occasioned by falling aggregate demand and employment, would increase the preference for liquid assets across the liquidity spectrum ranging from money or near-money (the most liquid), through short-term fixed interest securities through to long-term fixed interest securities and equities and, ultimately, physical plant and equipment (the least liquid of assets).

In formal terms, the uncertainty responsible for this phenomenon of liquidity preference can be represented by decision-making techniques based on multiple priors, sub-additive distributions, or fuzzy measure theory (Juniper, 2005). Let us take the first of these formalisms, incorporated into contemporary models of risk-sensitive control in systems characterised by a stochastic uncertainty constraint (measuring the gap between free and bound entropy) accounting for some composite of observation error, external perturbations, and model uncertainty. While the stochastic uncertainty constraint can be interpreted in ontological terms as one representing currently unknown but potentially knowable information (i.e. ambiguity), it can also be interpreted in terms of information that could never be known (i.e. fundamental uncertainty). For Keynes, calculations of expected returns were mere “conventions” designed to calm our disquietude, but they could never remove uncertainty by converting it into certainty equivalents.

Another source of both short-run and long-run departure from equilibrium has been described in Hyman Minsky’s (1992) analysis of Financial Instability, which was heavily influenced by both Keynes Michal Kalecki. As the economy began to recover from a period of crisis or instability, Minsky argued that endogenous forces would come into play that would eventually drive the system back into crisis. Stability would gradually be transformed into instability and crisis. On the return to a stable expansion path, after firms and households had repaired their balance-sheet structures, financial fragility would begin to increase as agents steadily came to rely more on external sources of finance, as firms began to defer the break-even times of their investment projects, and as overall levels of diversification in the economy steadily came to be eroded (see Barwell and Burrows, 2011, for an influential Bank of England study of Minskyian financial instability).  Minsky saw securitization (e.g. in the form of collateralized debt obligations etc.) as an additional source of fragility due to its corrosive effects on the underwriting system (effects that could never be entirely tamed through a resort to credit default swaps or more sophisticated hedging procedures). For Minsky, conditions of fragility, established preceding and during a crisis may only be partially overcome in the recovery stage, thus becoming responsible for ever deeper (hysteric) crises in the future[5].

An additional, perhaps more fundamental, reason for long-run instability is revealed by Piero Sraffa’s (1960) insights into the structural nature of shifts in the patterns of accumulation, within a multisectoral economy, as embodied in the notion of an invariant standard of value. Sraffa interprets David Ricardo’s quest for a standard commodity—one whose value would not change when the distribution of income between wages and profits was allowed to vary—as a quest that was ultimately self-defeating. This is because any standard commodity would have to be formally constructed with weights determined by the eigenvalue-structure of the input-output matrix. Nevertheless, changes in income distribution would lead to shifts in the composition of demand that, in turn would induce increasing or decreasing returns to scale. This would feed back onto the eigen-value structure of the input-output matrix, in turn requiring the calculation of another standard commodity (see Andrews, 2015, and Martins, 2019, for interpretations of Sraffa advanced along these lines). If we return to the case of the neoclassical growth model, Sraffa’s contribution to the debates in capital theory has completely undermined any notion of an optimal or “natural rate of interest” (Sraffa, 1960; Burmeister, 2000). From a policy perspective, this justifies an “anchoring” role for government policy interventions which aim to provide for both stability and greater equity in regard to both the minimum-wage (as an anchor for wage relativities) and determination of the overnight or ‘target’ rate of interest (as an anchor for relative rates-of-return).

From a modelling perspective, Martins (2019) insists that Sraffa drew a sharp distinction between a notion of ‘logical’ time (which is of relevance to the determination of “reproduction prices” on the basis of the labour theory of value, on the basis of a “snapshot” characterization of current input-output relations) and it’s counterpart, historical time (which is of relevance to the determination of social norms such as the subsistence wage, or policies of dividend-retention). When constructing stock-flow-consistent macroeconomic model this same distinction carries over to the historical determination of key stock-flow norms, which govern long-run behaviour in the model. Of course, in a long-run macroeconomic setting, fiscal and monetary policy interventions are also crucial inputs into the calculation of benchmark rates of accumulation (a feature which serves to distinguish these Post-Keynesian models from their neoclassical counterparts).[6]

Machine Learning and Fixed-point Theorems

In this paper’s discussion of macroeconomic phenomena, I have chosen to focus heavily on the determinants of movements away from stable, unique equilibria, in both the short-run and the long-run. Notions of equilibrium are central to issues of effectiveness in both econometrics and machine-learning. Of pertinence to the former, is the technique of cointegration and error-correction modelling. While the cointegrating vector represents a long-equilibrium, the error-correction process represents adjustment towards this equilibrium.  In a machine-learning context, presumptions of equilibrium underpin a variety of fixed-point theorems that play a crucial role in: (i) techniques of data reduction; (ii) efforts to eliminate redundancy within the network itself with the ultimate aim of overcoming the infamous “curse of dimensionality”, while preserving “richness of interaction”; and, (iii) the optimal tuning of parameters (and hyper-parameters that govern the overall model architecture). Specific techniques of data compression, such as Randomized Numerical Linear Algebra (Drineas and Mahoney, 2017), rely on mathematical techniques such as Moore-Penrose inverses and Tikhanov regularization theory (Barata and Hussein, 2011). Notions of optimization are a critical element in the application of these techniques. This applies, especially, to the gradient descent algorithms that are deployed for the tuning of parameters (and sometimes hyper-parameters) within the neural network. Techniques of tensor contraction and singular value decomposition are also drawn upon for dimensionality reduction is complex tensor networks (Cichoki et al., 2016, 2017). Wherever and whenever optimization techniques are required, some kind of fixed-point theorem comes into play. The relationship between fixed-point theorems, asymptotic theory, and notions of equilibrium in complex systems is not straightforward. See both Prokopenko et al., 2019 and Yanofsky, 2003, for a wide-ranging discussion of this issue, which opens onto a discussion of many inter-related “paradoxes of self-referentiality”.

For example, a highly-specialized literature on neural tangent kernels focuses on how kernel-based techniques can be applied in a machine learning context, to ensure that local rather than global maxima or minima are avoided during the whole process of gradient descent (see Yang, 2019). Here, the invariant characteristics of the kernel guarantee that tuning would satisfy certain robustness properties. An associated body of research on the tuning of parameters at the “edge of chaos”, highlights the importance of applying optimization algorithms close to the boundary of, but never within the chaotic region of dynamic flow (see Bietti and Mairal 2019, and Bertschinger and Natschläger, 2004). There are subtle formal linkages between the properties of neural tangent kernels and notions of optimization at the edge-of-chaos that I am unable to do justice to in this paper.

From a Post Keynesian perspective and despite this evolution in our understanding of optimization in a machine learning context, it would seem that efforts to apply the existing panoply of deep learning techniques may be thwarted by contrariwise aspects of the behaviour of dynamic macroeconomic system. For macroeconomists working with Real Business Cycle Models and their derivatives, none of this is seen as a problem because unreasonably-behaved dynamics are usually precluded by assumption. Although perturbations are seen to drive the business cycle in these models, agents are assumed to make optimal use of information, in the full knowledge of how the economy operates, so that government interventions simply pull the economy further away from equilibrium by adding more noise to the system. Although more recent dynamic stochastic general equilibrium (DSGE) models allow for various forms of market failure, notions of long-run equilibrium still play a fundamental role[7]. Instead, in a more realistic, Post Keynesian world, optimization algorithms would have to work very hard in their pursuit of what amounts to a “will-o-the-wisp”: namely, a system characterised by processes of shifting and non-stationary (hysteretic) equilibria[8].

Differential Programming

Recent discussions of machine learning and AI, have emphasized the significance of developments in differential programming. Yann LeCun (2018), one of the major contributors to the new Deep learning paradigm has noted that,

An increasingly large number of people are defining the networks procedurally in a data-dependent way (with loops and conditionals), allowing them to change dynamically as a function of the input data fed to them. It’s really very much like a regular program, except it’s parameterized, automatically differentiated, and trainable/optimizable.

One way of understanding this approach is to think of something that is a cross between a dynamic network of nodes and edges and a spread sheet. Each node contains a variety of functional formulas that draw on the inputs from other nodes and provides outputs that in turn, either feed into other nodes or can be observed by scopes. However, techniques of backpropagation and automatic differentiation can be applied to the entire network (using the chain rule while unfurling each of the paths in the network on the basis of Taylors series representations of each formula). This capability promises to overcome the limitations of econometric techniques when it comes to the estimation of large-scale models. For example, techniques of structural vector autoregression, which are multivariate extensions to univariate error-correction modelling techniques can only be applied to highly parsimonious, small-scale systems of equations.

Based on the initial work of Ehrhard and Regnier (2003), a flurry of research papers now deal with extensions to functional programming techniques to account for partial derivatives (Plotkin, 2020), higher-order differentiation and tensor calculus on manifolds (Cruttwell, Gallagher, & MacAdam, 2019), how best to account for computational effects (which are described in Rivas, 2018), and industrial-scale software engineering (The Statebox Team, 2019). Members of the functional programming and applied category theory community have drawn on the notion of a lens, as means for accommodating the bidirectional[9] nature of backpropagation[10] (Clarke et al., 2020; Spivak, 2019; Fong, Spivak and Tuyéras, 2017).


The potential flexibility and power of differential programming, could usher in a new era of policy-driven modelling, by allowing researchers to combine (i) traditionally aggregative macroeconomic models with multi-sectoral models of price and output determination (e.g. stock-flow-consistent Post Keynesian models and Sraffian or Marxian models of inter-sectoral production relationships); discrete-time and continuous-time models (i.e. hybrid systems represented integro-differential equations), and both linear and non-linear dynamics. This would clearly support efforts to develop more realistic models of economic phenomena.

The development of network-based models of dynamic systems has been given impetus by research in three main domains: brain science imaging, quantum tensor networks, and Geographical Information Systems in each case, tensor analysis of multiple-input and multiple-output nodes has played a key role. In each of these cases, the complexity associated with tensor algebra has been ameliorated by the deployment of diagrammatic techniques based on the respective use of Markov-Penrose’ diagrams, the diagrammatic Z-X calculus, and the development of “region-” rather than “point”-based topologies and mereologies. These same diagrammatic techniques have been taken up by the Applied Category Theory community to achieve both a deeper and more user-friendly understanding of lenses and other optics (Boisseau, 2020; Riley, 2018), alongside diagrammatic approaches to simply-typed, differential, and integral, versions of the lambda calculus (Lemay, 2017, Zeilberger and Giorgetti, 2015).

As I have argued, in more general terms, in Juniper (2018), the development of new software platforms based on diagrammatic reasoning could mean that differential programming techniques could potentially be disseminated to a much larger number of users who might have limited programming knowledge or skill (to some extent, today’s spreadsheets provide an example of this)[11]. In the case of AI, this could allow workers to regain control over machines which had previously either operated “behind their backs” or else, on the basis of highly specialized expertise. Improvements of this kind also have the potential to support higher levels of collaboration in innovation at the point-of-production. In the more restricted macroeconomic context, modelling could become less of a “black-box” and more of an “art” than a mystifying “science”. Diagrammatic approaches to modelling could help to make all of this more transparent. Of course, there are a lot of “coulds” in this paragraph. The development and use of technology can and should never be discussed in isolation form its political and organizational context. To a large extent, this political insight, was one of the main drivers and motivating forces for this paper.


[1] One intuitive way of thinking about this is that it would extend principles of “human centred manufacturing” into some of the more computational elements of the digital economy.

[2] See Christopher Olah’s blog entry for a helpful overview of various deep-learning architectures.

[3] For this reason, I will avoid any further discussion of convolution-based techniques and kernel methods, which have contributed, respectively, to rapid progress in image-classification and in applications of support-vector machines. An animated introduction to convolution-based techniques is provided by Cornellis (2018) while kernel-based techniques and the famous “kernel trick” deployed in support vector machines is lucidly described in Wright (2018). Rectified Linear Units or ReLU’s—the activation functions most commonly-used in deep learning neural networks—are examined in Brownlee (2019).

[4] The importance of symmetries in mathematical physics is examined in a recent paper by John Baez (2020), who investigates the source of symmetries in relation to Noether’s theorem.

[5] Some of these components of fragility, such as loss of diversification and deferment of breakeven times, would obviously be hard to capture in a highly aggregative macroeconomic model, but certain proxies could be constructed to this end.

[6] Of course, the rate at which labour—dead and living—is pulled out of production, also determines intra- and inter-sectoral economic performance, growth in trade, and overall rates of accumulation. It is also one of the key drivers of fundamental uncertainty for investors.

[7] See Stiglitz (2018) for a critical review of DSGE models, and Andrle and Solmaz (2017) for an empirical analysis of the business cycle, which raises doubts about the dynamic assumptions implied by a variety of macroeconomic models. The contribution of non-discretionary expenditure to instability in the business cycle has been highlighted by the recent Post Keynesian theoretical literature on the so-called “Sraffa super-multiplier” (Fiebiger, 2017; Fiebiger and Lavoie, 2017).

[8] Important sources of hysteresis, additional to those of a Minskyian nature, include those associated with rising unemployment, with its obvious impacts on physical and mental health, crime rates, and scarring in the eyes of prospective employers. Rates of innovation (and thus, productivity growth) are also adversely affected by declining levels of aggregate demand.

[9] The implementation function takes the vector of parameters and inputs and transforms them into outputs, while the request function takes parameters, inputs and outputs and emits a new set of inputs, whereas the update function takes parameters, inputs and outputs and transforms them into a new set of parameter values. Together, the update and request functions perform gradient descent with the request function passing back the inverted value of the gradient of total error with respect to the input. Each parameter is updated so that it moves a given step-size in the direction that most reduces the specified total error function

[10] For an introduction to some of the mathematical and programming-based techniques required for working with optics see Loregian (2019), Boisseau and Gibbons (2018), Culbertson and Kurtz (2013), and Román (2019).

[11] Software suites such as AlgebraicJulia and Statebox can already recognise the role of different types of string diagrams in representing networks, dynamical systems, and (in the latter case) commercial processes and transactions.


Anderson, C. (2008). The end of theory: The data deluge makes the scientific method obsolete. Wired, 23 June. Available at: http://www.wired.com/science/discoveries/magazine/16-07/pb_theory  (accessed 18 July, 2019).

Andrews, David (2015) . Natural price and the long run: Alfred Marshall’s misreading of Adam Smith. Cambridge Journal of Economics39: 265–279.

Andrle, Michal, Jan Brůha, Serhat Solmaz (2017). On the sources of business cycles: implications for DSGE models. ECB Working Paper, No 2058, May.

Baez, John (2020). Getting to the Bottom of Noether’s Theorem. arXiv:2006.14741v1 [math-ph] 26 Jun 2020.

Barata, J. C. A. & M. S. Hussein (2011). The Moore-Penrose Pseudoinverse. A Tutorial Review of the Theory. arXiv:1110.6882v1 [math-ph] 31 Oct 2011.

Barwell, R., & Burrows, O. (2011). Growing fragilities? Balance sheets in The Great Moderation. Financial Stability Paper No. 10, Bank of England.

Bengio, Yoshua; Aaron Courville; and Pascal Vincent (2014). Representation Learning: A Review and New Perspectives. arXiv:1206.5538v3 [cs.LG] 23 Apr 2014.

Bertschinger, N. & T. Natschläger (2004). Real-Time Computation at the Edge of Chaos in Recurrent Neural Networks. Neural Computation, July, 16(7): 1413-36.

Bietti, Alberto and Julien Mairal (2019). On the Inductive Bias of Neural Tangent Kernels. HAL Archive. https://hal.inria.fr/hal-02144221 (accessed 18 July, 2019)

Boisseau, Guillaume and Jeremy Gibbons (2018). What you needa know about yoneda: Profunctor optics and the yoneda lemma (functional pearl). Proc. ACM Program. Lang., 2(ICFP):84:1–84:27, July 2018.

Boisseau, Guillaume (2020). String diagrams for optics, arXiv:2002.11480v1 [math.CT] 11 Feb 2020.

Brownlee, J. (2019). A Gentle Introduction to the Rectified Linear Unit (ReLU) for Deep Learning Neural Networks. 9 Jan in Better Deep Learninghttps://machinelearningmastery.com/category/better-deep-learning/ . 

Burmeister, Edwin (2000) The Capital Theory Controversy. Critical Essays on Piero Sraffa’s Legacy in Economics, edited by Heinz D. Kurz. Cambridge: Cambridge University Press.

Carr, Nicholas (2010). The Shallows: How the Internet Is Changing the Way We Think, Read and Remember. New York: W.W. Norton and Company Inc.

Cichocki, Andrzej; Namgil Lee; Ivan Oseledets; Anh-Huy Phan; Qibin Zhao; and Danilo P. Mandic (2016). Tensor Networks for Dimensionality Reduction and Large-scale Optimization: Part 1 Low-Rank Tensor Decompositions. Foundations and Trends in Machine Learning. 9(4-5), 249-429.

Cichocki, Andrzej ; Anh-Huy Phan; Qibin Zhao; Namgil Lee; Ivan Oseledets; Masashi Sugiyama; and Danilo P. Mandic (2017). Tensor Networks for Dimensionality Reduction and Large-scale Optimization: Part 2 Applications and Future Perspectives. Foundations and Trends in Machine Learning. 9(6), 431-673.

Clarke, B., D. Elkins, J. Gibbons, F. Loregian, B. Milewski, E. Pillore, & M. Roman (2020). Profunctor Optics, a Categorical Update. arXiv:2001.07488v1 [cs.PL] 21 Jan 2020.

Cornelisse, Daphne (2018). “An intuitive guide to Convolutional Neural Networks”, available at FreeCodeCamp, https://www.freecodecamp.org/news/an-intuitive-guide-to-convolutional-neural-networks-260c2de0a050/ .

Cruttwell, Gallagher, & MacAdam (2019). Towards formalizing and extending differential programming using tangent categories. Extended Abstract, Proc. ACT 2019, available at: http://www.cs.ox.ac.uk/ACT2019/preproceedings/Jonathan%20Gallagher,%20Geoff%20Cruttwell%20and%20Ben%20MacAdam.pdf .

Culbertson, J. & K. Sturtz (2013). Bayesian Machine Learning via Category Theory. arXiv:1312.1445v1 [math.CT] 5 Dec2013.

Ehrhard, Thomas and Laurent Regnier (2003). The differential lambda calculus. Theoretical Computer Science, 309 (1-3):1-41.

Drineas, Petros and Michael W. Mahoney (2017). Lectures on Randomized Numerical Linear Algebra. arXiv:1712.08880v1 [cs.DS] 24 Dec 2017.

Fiebiger, B. (2017). Semi-autonomous household expenditures as the causa causans of postwar US business cycles: the stability and instability of Luxemburg-type external markets. Cambridge Journal of Economics, vol. 42, Issue 1, 2018, pp. 155–175.

Fiebiger, B., & Lavoie, M. (2017). Trend and business cycles with external markets: Non-capacity generating semi-autonomous expenditures and effective demand. Metroeconomica.2017;00:1–16.

Fong, Brendan, David Spivak and Rémy Tuyéras’s (2017). Backpropagation as Functor: A compositional perspective on supervised learning.  https://arxiv.org/abs/1711.10455v3.

Gershenfeld, Neil, Alan Gershenfeld, and Joel Cutcher-Gershenfeld (2018). Designing Reality: How to Survive and Thrive in the Third Digital Revolution . New York: Basic Books.

Hedges Jules, Jelle Herold (2019). Foundations of brick diagrams. rXiv:1908.10660v1 [math.CT] 28 Aug 2019.

Juniper, J. (2018). Economic Philosophy of the Internet-of-Things. London: Routledge.

Juniper, J. (2005). A Keynesian Critique of Recent Applications of Risk-Sensitive Control Theory in Macroeconomics, Contemporary Post Keynesian Analysis, proceedings of the 7th International Post Keynesian Workshop, Northhampton: Edward Elgar, UK.  

Keynes, J. M. (1936). The General Theory of Employment, Interest and Money, London, Macmillan, Retrieved from: http://www.hetwebsite.net/het/texts/keynes/gt/gtcont.htm .

Lin, H. W., M. Tegmark & D. Rodnick (2017). Why does deep and cheap learning work so well? J. of Stat. Physics. arXiv:1608.08225v4 [cond-mat.dis-nn] 3 Aug 2017.

LeCun, Yann (2018). Deep Learning est mort. Vive Differentiable Programming! Facebook blog entry, January 6, 2018:  https://www.facebook.com/yann.lecun/posts/10155003011462143 020-01-07

Lemay Jean-Simon Pacaud (2017). Integral Categories and Calculus Categories. Master of Science Thesis, University of Calgary, Alberta.

Loregian, Fosco (2019). Coend calculus—the book formerly known as ‘This is the co/end’. arXiv:1501.02503v5 [math.CT] 21 Dec 2019.

Lovelock, James (2019). Novacene: The Coming Age of Hyperintelligence. London: Allen Lane.

Martins, Nuno Ornelas (2019). The Sraffian Methodenstreit and the revolution in economic theory. Cambridge Journal of Economics, 43: 507–525.

Minsky, Hyman P. (May 1992). The Financial Instability Hypothesis. The Jerome Levy Economics Institute of Bard College, Working Paper No. 74: 6–8. http://www.levy.org/pubs/wp74.pdf .

Olah, Christopher (2015). Colah, Blog entry on “Neural Networks, Types, and Functional Programming”. Posted on September 3, http://colah.github.io/posts/2015-09-NN-Types-FP/ .

Plotkin, Gordon (2020). A complete axiomatisation of partial differentiation. The Spring Applied Category Theory Seminar at University of California, Riverside, 7 June, 2020,   http://math.ucr.edu/home/baez/ACT@UCR/index.html#plotkin .

Poggio, T., H. Mhaskar, L. Rosasco, B. Miranda & Q. Liao (2017). Why and When Can Deep—but not Shallow—Networks Avoid the Curse of Dimensionality: A Review. International Journal of Automation and Computing, 14(5), October 2017, 503-519.

Prokopenko, Harre, Lizier, Boschetti, Peppas, Kauffman (2019). Self-referential basis of undecidable dynamics: from the Liar paradox and The Halting Problem to The Edge of Chaos. arXiv:1711.02456v2 [cs.LO] 21 Mar 2019.

Riley, M. (2018). Categories of Optics. arXiv:1809.00738v2 [math.CT] 7 Sep 2018.

Rivas, E. (2018). Relating Idioms, Arrows and Monads from Monoidal Adjunctions. Chapter in R. Atkey and S. Lindley (Eds.): Mathematically Structured Functional Programming (MSFP 2018) EPTCS 275, 2018, pp. 18–33.

Román, Mario (2019). Profunctor optics and traversals. MSc Thesis in Mathematics and Foundations of Computer Science, Trinity, Oxford University. arXiv:2001.08045v1 [cs.PL] 22 Jan 2020.

Spivak, David I. (2019). Generalized Lens Categories via Functors Cop → Cat. arXiv:1908.02202v2 [math.CT] 7 Aug 2019.

Sraffa, Piero (1960) Production of Commodities by means of Commodities: A Prelude to the Critique of Neo-Classical Economics. Cambridge: Cambridge University Press.

Tegmark, Max (2017). Life 3.0: Being Human in the Age of Artificial Intelligence. London: Penguin Books.

The Statebox Team (2019). The Mathematical Specification of the Statebox Language, Version June 27, 2019, https://statebox.org/research/ .

Stiglitz, J. E., (2018) Where modern macroeconomics went wrong, Oxford Review of Economic Policy, 34(1-2), pp. 70–106.

Wright, A. (?). Appendix A-Brief Introduction to Kernels. Mimeo. University of Lancaster. https://www.lancaster.ac.uk/pg/wrighta3/STOR603_Appendix_A.pdf .

Yang, G. (2019). Scaling Limits of Wide Neural Networks with Weight Sharing: Gaussian process behavior, gradient independence, and neural tangent kernel derivation. arXiv preprint arXiv:1902.04760, 2019.

Yanofsky (2003). A universal approach to self-referential paradoxes, incompleteness and fixed-points. arXiv:math/0305282v1 [math.LO] 19 May 2003.

Zeilberger, Noam and Alain Giorgetti (2015). A correspondence between rooted planar maps and normal planar lambda terms. Logical Methods in Computer Science, Vol. 11, 3(22): 1–39.

Zuboff, Shoshana (2019).  The Age of Surveillance Capitalism: The Fight for a Human Future at the New Frontier of Power. London: Profile Books.

Semantic Technologies for Disaster Management: Network Models and Methods of Diagrammatic Reasoning


The Chapter will provide a brief and informal introduction to diagrammatic reasoning (DR) and network modelling (NM) using string diagrams, which can be shown to possess the same degree of rigor as symbolic algebra, while achieving greater abbreviative power (and  pedagogical insight) than more conventional techniques of diagram-chasing. This review of the research literature will set the context for a detailed examination of two case-studies of semantic technologies which have been applied to the management of emergency services and search-and-rescue operations. The next section of the Chapter will consider the implications of contemporary and closely related developments in software engineering for disaster management. Conclusions will follow.


This Chapter is concerned with developments in applied mathematics and theoretical computing that can provide a formal and technical support for practices of disaster management. To this end it will draw on recent developments in applied category theory , which inform semantic technologies. In the interests of brevity, it will be obliged to eschew formal exposition of these techniques, but to this end, comprehensive references will be provided. The justification for what might at first seem to be an unduly narrow focus, is that applied category theory facilitates translation between different mathematical, computational and scientific domains.

For its part, Semantic Technology (ST) can be loosely conceived as an approach treating the World-Wide-Web as a “giant global graph”, so that valuable and timely information can be extracted from it using rich structured-query languages and extended description logics. These query languages must be congruent with pertinent (organizational, application, and database) ontologies so that the extracted information can be converted into intelligence. Significantly, database instances can extend beyond relational or graph databases, to include Boolean matrices, relational data embedded within the category of linear relations, and that pertaining to systems of differential equations in finite vector space, or even quantum tensor networks within a finite Hilbert space.

More specifically, this chapter will introduce the formalism of string diagrams, which were initially derived from the work of the mathematical physicists, Roger Penrose (1971) and Richard Feynman (1948). However, this diagrammatic approach has since been extended and re-interpreted  by category theorists such as Andre Joyal and Roy Street (1988, 1991). For example, Feynman diagrams can be viewed as morphisms in the category Hilb of Hilbert spaces and bounded linear operators (Westrich, 2006, fn. 3: 8), while Baez and Lauda (2009) interpret them as “a notation for intertwining operators between positive-energy representations of the Poincaré group”. Penrose diagrams can be viewed as a representation of operations within a tensor category.

Joyal and Street have demonstrated that when these string diagrams are manipulated in accordance with certain axioms—the latter taking the form of a set of equivalence relations established between related pairs of diagrams—the movements from one diagram to another can be shown to reproduce the algebraic steps of a non-diagrammatic proof. Furthermore, they can be shown to possess a greater degree of abbreviative power. This renders an approach using string diagrams extremely useful for teaching, experimentation, and exposition.

In addition to these conceptual and pedagogical advantages, however, there are additional implementation advantages associated with string diagrams including: (i) those of compositionality and layering (e.g. in Willems’s 2007  behavioural approach to systems theory, complex systems can be construed as the composites of smaller and simpler building blocks, which are then linked together in accordance with certain coherence conditions); (ii) a capacity for direct translation into functional programming (and thus, into propositions within a linear or resource-using logic); and, (iii) the potential for the subsequent application of software design and verification tools. It should be appreciated that these formal attributes will become increasingly important as the correlative features of what some have described as the digital economy.

This chapter will consider the specific role of string diagrams in the development and deployment of semantic technologies, which in turn have been developed for applications of relevance to disaster management practices. Techniques based on string diagrams have been developed to encompass a wide variety of dynamic systems and application domains, such as Petri nets, the π-calculus, and Bigraphs (Milner, 2009), Bayesian networks (Kissinger & Uijlen, 2017), thermodynamic networks (Baez and Pollard, 2017), and quantum tensor networks (Biamonte & Bergholm, 2017), as well as reaction-diffusion systems (Baez and Biamonte, 2012). Furthermore, they have the capacity to encompass graphical forms of linear algebra (Sobociński, Blog), universal algebras (Baez, 2006), and signal flow graphs (Bonchi, Sobociński and Zanasi (2014, 2015), along with computational logics based on linear logic and graph rewriting (on this see Mellies, 2018; and Fong and Spivak, 2018, for additional references).

1.  Applied Category Theory

Category theory and topos theory have taken over large swathes in the field of formal or theoretical computation, because categories serve to link together the structures found in algebraic topology, and with the logical connectives and inferences to be found in formal logic, as well as with recursive processes and other operations in computation. The following diagram taken from Baez and Stay (2011), highlights this capability.

John Bell (1988: 236) succinctly explains why it is that category theory also possesses enormous ormous powers of generalization:

A category may be said to bear the same relation to abstract algebra as does the latter to elementary algebra. Elementary algebra results from the replacement of constant quantities (i.e. numbers) by variables, keeping the operations on these quantities fixed. Abstract algebra, in its turn, carries this a stage further by allowing the operations to vary while ensuring that the resulting mathematical structures (groups, rings, etc) remain of a prescribed kind. Finally, category theory allows even the kind of structure to vary: it is concerned with structure in general.

Category theory can also be interpreted as a universal approach to the analysis of process, across various domains including: (a) mathematic practice (theorem proving); (b) physical systems (their evolution and measurement); (c) computing (data types and programs); (d) chemistry (chemicals and reactions); (e) finance (currencies and various transactions); (f) engineering (flows of materials and production).

This way of thinking about processes now serves as a unifying interdisciplinary framework that researchers within business and the social sciences have also taken up. Alternative approaches to those predicated on optimizing behaviour on the part of individual economic agents include the work evolutionary economists and those in the business world who are obliged to work with computational systems designed for the operational management of commercial systems. However, these techniques are also grounded in conceptions of process

Another way of thinking about dynamic processes is in terms of circuit diagrams, which can represent displacement, flow, momentum and effort—phenomenon modelled by the Hamiltonians and Lagrangians of Classical Mechanics. It can be appreciated that key features of economic systems are also amenable to diagrammatic representations of this kind, including asset pricing based on notion of arbitrage, a concept initially formalized by Augustin Cournot in 1838. Cournot’s analysis arbitrage conditions is grounded in Kirchoff voltage law (Ellerman, 1984). The analogs of displacement, flow, momentum and effort are depicted below for a wide range of disciplines.

Applied Category Theory: in the US, contemporary developments in applied category theory (ACT) have been spurred along and supported by a raft of EU, DARPA and ONR Grants. A key resource on ACT is Fong and Spivak’s (2018) downloadable text on compositionality. This publication explores the relationship between wiring diagrams or string diagrams and a wide variety of mathematical and categorical constructs, including as a means for representing symmetric monoidal preorders, signal flow graphs, along with functorial translation between signal flow graphs and matrices and other aspects of functorial semantics, graphical linear algebra, hypergraph categories and operads, applied to electric circuits and network compositionality. Topos theory is introduced to characterise the logic of system behaviour on the basis of indexed sets, glueings, and sheaf conditions for every open cover.

2. Diagrammatic Reasoning

Authors such as Sáenz-Ludlow and Kadunz (2015), Shin (1995), Sowa (2000), and Stjernfelt (2007), who have published research on knowledge representation and diagrammatic approaches to reasoning, tend to work within a philosophical trajectory that stretches from F. W. Schelling and C. S. Peirce, through to E. Husserl and A. N. Whitehead, then on to M. Merleau-Ponty and T. Adorno. Where Kant and Hegel privileged symbolic reasoning over the iconic or diagrammatic, Peirce, Whitehead, and Merleau-Ponty followed the lead of Schelling for whom ‘aesthetics trumps epistemology’! It is, in fact, this shared philosophical allegiance that not only links diagrammatic research to the semantic (or embodied) cognition movement (Stjernfeld himself refers to the embodied cognition theorists Eleanor Rosch, George Lakoff, Mark Johnson, Leonard Talmy, Mark Turner, and Gilles Fauconnier), but also to those researchers who have focused on issues of educational equity in the teaching of mathematics and computer science, including Ethnomathematics and critical work on ‘Orientalism’ specialized to emphasize a purported division between the ‘West and the Rest’ in regard to mathematical and computational thought and practice.

As such, insights from this research carry over to questions of ethnic ‘marginalization’ or ‘positioning’ in the mathematical sciences (see the papers reproduced in Forgasz and Rivera, eds., 2012 and Herbel-Eisenmann et al., 2012). In a nutshell, diagrammatic reasoning is sensitive to both context and positioning and, thus, is closely allied to this critical axis of mathematics education.

The following illustration of the elements and flows associated with diagrammatic forms of reasoning comes from Michael Hoffman’s (2011) explication of the concept first outlined by the American philosopher and logician, Charles Sanders Peirce.

The above Figure depicts three stages in the process of diagrammatic reasoning: (i) constructing a diagram as a consistent representation of key relations; (ii) analysing a problem on the basis of this representation; and (iii) experimenting with the diagram and then observing the results. Consistency is ensured in two ways. First, the researcher or research team develop an ontology specifying elements of the problem and the relations holding between these elements, along with pertinent rules of operation. Second, language is specified in terms of both syntactical and semantic properties. Furthermore, in association with this language, a rigorous axiomatic system is specified, which both constrains and enables any pertinent diagrammatic transformations.

3a. Case-Study One:

A 2010 paper by SAP Professors, Paulheim and Probst reviews an application of STs to the management and coordination of emergency services in the Darmstadt region of Germany. The aim of the following diagram, reproduced from their work, is to highlight the fact that, from a computational perspective, the integrative effort of STs can apply to different organizational levels: that of the common user interface, shared business logics and that of data sources.

In their software engineering application, the upper-level ontology DOLCE is deployed to link a core domain ontology together with a user-interface interaction ontology. In turn, each of these ontologies draws on inputs from an ontology on deployment regulations and various application ontologies. Improved search capabilities across this hierarchy of computational ontologies, are achieved through the adoption of the ONTOBROKER and F-Logic systems.

3b. Case-study Two:

An important contribution to the field of network modelling has come from the DARPA-funded CASCADE Project (Complex Adaptive System Composition and Design Environment), which has invested in long-term research into the “system-of-systems” perspective (see John Baez’s extended discussion of this project on his Azimuth blog). This research has been influenced by Willems’s (2007) behavioural approach to systems, which in turn, is based on the notion that large and complex systems can be built up from simple building blocks.

Baez et al. (2020) introduce ‘network models’ to encode different ways of combining networks both through overlaying one model on top of another and by setting each model side by side. In this way, complex networks can be constructed using simple networks as components. Vertices in the network represent fixed or moving agents, while edges represent communication channels.

The components of their networks are constructed using coloured operads, which include vertices representing entities of various types and edges representing the relationships between these entities. Each network model gives rise to a typed operad with an associated canonical algebra, whose operations represent ways of assembling a more complex network from smaller parts. The various different ways to compose these operations characterize a more general notion of an operation, which must be complemented by ways of permuting the arguments of an operation a process yielding a permutation group of inputs and outputs).

In research conducted under the auspices of the CASCADE Project, Baez, Foley, Moeller, and Pollard (2020) have worked out how to combine two formalisms. First, there are Petri nets, commonoly used as an alternative to process algebras as a foralism for business process management. The vertices in a Petri net represent collections of different types of entities (species) with morphisms between them used to describe processes (transitions) that can be carried out by combining various sets of entities (conceived as resources or inputs into a transition node or process of production) together to make new sets of entities (concived as outputs or vertices are positioned after the relevant transition node). The stocks of each type of entity that is available is enumerated as a ‘marking’ specific to each type or colour together with the set of outputs that can be produced by activated the said transition.

Second, there are network models, which describe processes that a given collection of agents (say, cars, boats, people, planes in a search-and-rescue operation) can carry out. However, in this kind of network, while each type of object or vertex can move around within a delineated space, they are not allowed to turn into other types of agent or object.

In these networks, morphisms are functors (generalised functions) which describe everything that can be done with a specific collection of agents. The following Figure depicts this kind of operational network in an informal manner, where icons represent helicopters, boats, victims floating in the sea, and transmission towers with communication thresholds.

By combining Petri nets with an underlying network model resource-using operations can be defined. For example, a helicopter may be able to drop supplies gathered from different depots and packaged into pallets, onto the deck of a sinking ship or to a remote village cut off by an earthquake or flood.

The formal mechanism for combining a network model with a Petri net relies on treating  different type of entities as catalysts, in the sense that the relevant species are neither increased nor decreased in number by any given transition. The derived category is symmetric monoidal and possesses a tensor product (representing processes for each catalyst that occur side-by-side), a coproduct (or disjoint union of amounts of each catalyst present), and within each subcategory of a particular catalyst, an internal tensor product describes how one process can follow another while reusing the pertinent catalysts.

The following diagram taken from Baez et al. (2020), illustrates the overlaying process which enables more complex networks to be constructed from simpler components. The use of the Grothendieck construction in this research ensures that when two or more diagrams are overlayed there will be no ‘double-counting’ of edges and vertices. When components are ‘tensored’ each of the relevant blocks would be juxtaposed “side-by-side”.

Each network model is characterized by a “plug-and-play” feature based on an algebraic component called an operad. The operad serves as the construct for a canonical algebra, whose operations are ways of assembling a network of the given kind from smaller parts. This canonical algebra, in turn, accommodates a set of types, a set of operations, ways to compose these operations to arrive at more general operations, and ways to permute an operation’s arguments (i.e. via a permutation group), along with a set of relevant distance constraints (e.g. pertinent communication thresholds for each type of entity) .

One of Baez’s co-authors, John Foley, works for Metron, Inc., VA, a company which specializes in applying the advanced mathematics of network models to such phenomena as “search-and-rescue” operations, the detection of network incursions, and sports analytics. Their 2017 paper mentions a number of formalisms that have relevance to “search-and-rescue” applications, especially the ability to distinguish between different communication channels (different radio frequencies and capacities) and vertices (e.g. planes, boats, walkers, individuals in need of rescue etc.) and the capacity to impose distance constraints over those agents who may fall outside the reach of communication networks.

In related research paper, Schultz, Spivak, Vasilakopoulou, Wisnesky (2016) argue thay dynamical systems can be gainfully thought of as ‘machines’ with inputs and outputs, carrying some sort of signal that occurs through some notion of time”. Special cases of this general approach include discrete, continuous, and hybrid dynamical systems. The authors deploy lax functors out of monoidal categories, which provide them with a language of compositionality. As with Baez and his co-authors, Schultz et al. (2016) draw on an operadic construct so as to understand systems that result from an “arbitrary interconnection of component subsystems”. They also draw on the mathematics of sheaf theory, to flexibly capture the crucial notion of time. The resulting sheaf-theoretic perspective relates continuous- and discrete-time systems together via functors (a kind of generalized ‘function of functions’, which preserves structure). Their approach can also account for synchronized continuous time, in which each moment is assigned a specific phase within the unit interval.

4. Related Developments in Software Engineering

This section of the Chapter examines contemporary advances in software engineering that have implications for ‘system-of-sytems’ approaches to semantic technology. The work of the Statebox group at the University of Oxford and that of Evan Patterson, from Stanford University, who is also affiliated with researchers from the MIT company, Categorical Informatics, will be discussed to indicate where these new developments are likely to be moving in the near future. This will be supplemented by an informal overview of some recent innovations in functional programming, which have been informed by the notion of a derivative applied to an algorithmic step. These initiatives have the potential to transform software for machine-learning and the optimization of networks

The Statebox team based at Oxford University have developed a language for software engineering that uses diagrammatic representations of generalized Petri nets. In this context, transitions in the net are morphisms between data-flow objects represent terminating functional programming algorithms. In Statebox (integer and semi-integer) Petri nets are constructed with both positive and negative tokens to account for contracting. Negative tokens represent borrowing while positive tokens represent lending and, likewise, the taking of short and long positions in asset markets. This allows for the representation of smart contracts, conceived as separable nets. Nets are also endowed with interfaces that allow for channelled communications through user-defined addresses. Furthermore, guarded and timed nets, with side-effects (which are mapped to standard nets using the Grothendieck construction), offer greater expressive power in regard to the conditional behaviour affecting transitions (The Statebox Team, 2018).

Patterson (2017) begins his paper with a discussion of description logics (e.g. OWL, WC3), which he interprets as calculi for knowledge representation (KR). These logics, which are the actual substrates responsible for the World-Wide-Web (WWW), lie somewhere between propositional logic and first-order predicate logic possessing the capability to express the (∃,∧,T,=) fragment of first-order logic. Patterson highlights the trade-off that must be made between computational tractability and expressivity before introducing a third knowledge representation formalism that interpolates between description logic and ontology logs (see Spivak and Kent, 2012, for an the extensive description of ologs, which express key constructs from category theory, such as products and coproducts, pullbacks and pushforwards, and representations of recursive operations using diagrams labelled with concepts drawn from everyday conversation). Patterson (2017) calls this construct the relational ontology log, or relational olog, because it is based on, Rel, the category of sets and relations and, as such, draws on relational algebra, which is the (∃,∧, , T,⊥,=) fragment of first-order logic. He calls Spivak and Kent’s, 2012, version, a functional olog to avoid any confusion, because these are solely based on Set, the category of sets and functions. Relational ologs achieve their expressivity through categorical limits and colimits (products, pullbacks, pushforwards, and so forth

The advantages of Patterson’s framework are that functors allow instance data to be associated with a computational ontology in a mathematically precise way, by interpreting it as a relational or graph database, Boolean matrix, or category of linear relations. Moreover, relational ologs are, by default, typed, which he suggests can mitigate the maintainability challenges posed by the open world semantics of description logic.

String diagrams (often labelled Markov-Penrose diagramsby those working in the field of brain science imaging) are routinely deployed by data-scientists used to represent the structure of deep-learning convolution neural networks. However, string diagrams can also serve as a tool for representing the computational aspects of machine-learning.

For example, influenced by the program idioms of machine-learning, Ghica and Muroya (2017) have developed what they choose to call a ‘Dynamic Geometry of Interaction Machine’, which can be defined as a state transition system operating whose transitions not only account for ‘token passing’ but also for ‘graph rewriting’ (where the latter can be construed as a graph-based approach to the proving of mathematical hypotheses and theories). Their proposes system is supported by diagrammatic implementation based on the proof structures of the multiplicative and exponential fragment of linear logic (MELL). In Muroya, Cheung and Ghica (2017), this logical approach is complemented by a sound call-by-value lambda calculus inspired, in turn by Peircean notions of abductive inference. The resulting bimodal programming model operates in both: (a) direct mode, with new inputs provided, new outputs obtained; and, (b) learning mode, with special inputs applied for which outputs are known; to achieve optimal tuning of parameters to ensure desired outputs approach actual outputs. The authors contend that their holistic approach is superior to that of the TensorFlow software package developed for machine-learning, which they describe as a ‘shallow embedding’ of a domain specific language (DSL) into PYTHON” rather than a ‘stand-alone’ programming language.

Adopting a somewhat different approach, Cruttwell, Gallagher and MacAdam (2019) extend Plotkin’s differential programming framework, which is itself a generalization of differential neural computers, where arbitrary programs with control structures encode smooth functions also represented as programs. Within this generalized domain, the derivative can be directly applied to programs or to algorithmic steps and, furthermore, can be rendered entirely congruent with categorical approaches to Riemannian and Differential geometry such as Lawvere’s Synthetic Differential Geometry.

Cruttwell and his colleagues go on to observe that, when working in a simple neural network, back-propagation takes the derivative of the error function, then uses the chain rule to push errors backwards. They point out that, for convolution neural networks, the necessary procedure is less straightforward due to the presence of looping constructs.

In this context, the authors further note that attempts to work with the usual ‘if-then-else’ and ‘while’ commands can also be problematic. To overcome these problems associated with recursion, they deploy what have been called ‘join restriction tangent categories’, which express the requisite domain of definition and detect and achieve disjointness of domains, while expressing iteration using the join of disjoint domains (i.e. in technical terms, this is the trace of a coproduct in the idempotent splitting). The final mathematical construct they arrive at, is that of a differential join restriction category along with the associated join restriction functor which, they suggest, admits a coherent interpretation of differential programming.

It should be stressed that each of these category-theoretic initiatives to formalize the differential of an algorithmic step will become important in future efforts to develop improved, yet diagrammatically-based forms of software for machine learning that have greater capability and efficiency than existing software suites. The fact that both differential and integral categories can be provided with a coherent string diagram formalism (Lemay, 2017) provides a link back to the earlier discussion about the role of diagrammatic reasoning in semantic technologies.

It is clear that techniques of this kind could also be applied to a wide variety of network models (e.g. for the centralized and decentralized control of hybrid cyber-physical systems), where optimization routines may be required (including those for effective disaster management).

5. Conclusion

In conclusion, the innovations in software engineering described above, have obvious implications for those attempting to  develop new semantic technologies for the effective management of emergency services and search-and-rescue operations in the aftermath of a major disaster. Hopefully, the material surveyed in this Chapter should serve to highlight the advantages of a category-theoretic approach to the issue at hand, along with the specific benefits of adopting an approach that is grounded in the pedagogical, computational, and formal representational power of string-diagrams, especially within a networked computational  environment charactrised by Big Data, parallel processing, hybridity, and some degree of decentralized control.

While a Chapter of this kind cannot go into too much detail about the formalisms that have been discussed, it is to be hoped that enough pertinent references have been provided for those who would like to find out more about the mathematical detail. Of course, it is not always necessary to be a computer programmer both to understand and to effectively deploy powerful suites of purpose-built software. It is also to be hoped that diagrammatic reasoning may assist the interested reader in acquiring a deeper understanding of the requisite mathematical techniques.

Author: Professor Dr. James Juniper – Conjoint Academic, University of Newcastle; PhD in Economics, University of Adelaide

Chapter References

Baez, John (2006). Course Notes on Universal Algebra and Diagrammatic Resoning. Date accessed 15/11/19. Available at http://math.ucr.edu/home/baez/universal/

Baez, John C. & Jacob D. Biamonte (2012). A Course on Quantum Techniques for Stochastic Mechanics. arXiv:1209.3632v1 [quant-ph] 17 Sep 2012.

Baez, John C., Brandon Coya and Franciscus Rebro (2018). Props in Network Theory. Theory and Applications of Categories, 33(25): 727-783.

Baez, J., J. Foley, J. Moeller, and B. Pollard (2020). Network Models. (accessed 1/7/2020)  arXiv:1711.00037v3  [math.CT]  27 Mar 2020.

Baez, John and Brendan Fong (2018). A Compositional Framework for Passive Linear Networks. arXiv:1504.05625v6  [math.CT]  16 Nov 2018

Baez, John C. & Aaron Lauda (2009). A Prehistory of n-Categorical Physics. Date accessed 5/02/2018. https://arxiv.org/abs/0908.2469.

Baez, John C. and Blake Pollard (2017). A compositional framework for reaction networks. Reviews in Mathematical Physics, 29 (2017), 1750028.

Baez, John C. and Michael Stay (2011). Physics, Topology, Logic and Computation: A Rosetta Stone. New Structures for Physics, ed. Bob Coecke, Lecture Notes in Physics vol. 813, Springer, Berlin, 95-174.

Bell J. T. (1998). A Primer of Infinitesimal Analysis, Cambridge, U.K. Cambridge University Press.

Biamonte, J. and V. Bergholm (2017). Quantum Tensor Networks in a Nutshell. Cornell University Archive. Date accessed 15/11/19. arXiv:1708.00006v1 [quant-ph] 31 Jul 2017.

Blinn, James F. (2002). Using Tensor diagrams to Represent and solve Geometric Problems. Microsoft Research, Publications, Jan. 1. Date accessed 15/11/19.  https://www.microsoft.com/en-us/research/publication/using-tensor-diagrams-to-represent-and-solve-geometric-problems/ .

Bonchi, F., P. Sobociński and F. Zanasi (2015). Full Abstraction for Signal Flow Graphs. In Principles of Programming Languages, POPL’15, 2015.

Bonchi, F., P. Sobociński and F. Zanasi (2014). A Categorical Semantics of Signal Flow Graphs. CONCUR 2014, Ens de Lyon.

Cichocki, Andrzej; Namgil Lee; Ivan Oseledets; Anh-Huy Phan; Qibin Zhao; and Danilo P. Mandic (2016). Tensor Networks for Dimensionality Reduction and Large-scale Optimization: Part 1 Low-Rank Tensor Decompositions. Foundations and Trends in Machine Learning. 9(4-5), 249-429.

Cichocki, Andrzej ; Anh-Huy Phan; Qibin Zhao; Namgil Lee; Ivan Oseledets; Masashi Sugiyama; and Danilo P. Mandic (2017). Tensor Networks for Dimensionality Reduction and Large-scale Optimization: Part 2 Applications and Future Perspectives. Foundations and Trends in Machine Learning. 9(6), 431-673.

Cruttwell, Gallagher & MacAdam (2019). Towards formulating and extending differential programming using tangent categories. Extended abstract, ACT 2019. Date accessed 15/11/19. Available at: http://www.cs.ox.ac.uk/ACT2019/preproceedings/Jonathan%20Gallagher,%20Geoff%20Cruttwell%20and%20Ben%20MacAdam.pdf .

Ehrhard T., and L. Regnier (2003). The differential lambda-calculus. Theoretical Computer Science. 309, 1–41.

Ellerman, David (2000). Towards an Arbitrage Interpretation of Optimization Theory. (accessed 1/7/20), http://www.ellerman.org/Davids-Stuff/Maths/Math.htm .

Feynman, R. P. (1948). “Space-time approach to nonrelativistic quantum mechanics,” Review of Modern Physics, 20, 367.

Fong, Brendan and David I. Spivak (2018). Seven Sketches in Compositionality:An Invitation to Applied Category Theory. Date accessed 15/11/19. Available at http://math.mit.edu/~dspivak/teaching/sp18/7Sketches.pdf .

Forgasz, Helen and Ferdinand Rivera (eds.) (2012). Towards Equity in Mathematics Education: Gender, Culture, and Diversity. Advances in Mathematics Education Series. Dordrecht, Heidelburg: Springer.

Herbel-Eisenmann, Beth, Jeffrey Choppin, David Wagner, David Pimm (eds.) (2012). Equity in Discourse for Mathematics Education Theories, Practices, and Policies. Mathematics Education Library, Vol. 55. Dordrecht, Heidelburg: Springer.

Hoffman, M. H. G. (2011). Cognitive conditions of diagrammatic reasoning. Semiotica, 186 (1/4), 189–212.

Joyal, A. and R. Street (1988). Planar diagrams and tensor algebra. Unpublished manuscript. Date accessed 15/11/19. Available from Ross Street’s website: http://maths.mq.edu.au/~street/.

Joyal, A. and R. Street (1991). The geometry of tensor calculus, I. Advances in Mathematics, 88, 55–112.

Kissinger, Aleks and Sander Uijlen (2017). A categorical semantics for causal structure. https://arxiv.org/abs/1701.04732v3 .

Lemay, Jean-Simon Pacaud (2017). Integral Categories and Calculus Categories. PhD Thesis, University of Calgary, Alberta.

Melliès, Paul-André (2018). Categorical Semantics of Linear Logic. Date accessed 15/11/19. Available at: https://www.irif.fr/~mellies/mpri/mpri-ens/biblio/categorical-semantics-of-linear-logic.pdf .

Milner, Robin (2009). The Space and Motion of Communicating Agents. Cambridge University Press.

Moeller, Joe & Christina Vasilakopolou (2019). Monoidal Grothendieck Construction. arXiv:1809.00727v2 [math.CT] 18 Feb 2019.

Muroya, Koko and Dan Ghica (2017). The Dynamic Geometry of Interaction Machine: A Call-by-need Graph Rewriter. arXiv:1703.10027v1 [cs.PL] 29 Mar 2017.

Muroya, Koko; Cheung, Steven and Dan R. Ghica (2017). Abductive functional programming, a semantic approach. arXiv:1710.03984v1 [cs.PL] 11 Oct 2017.

Patterson, Evan (2017). Knowledge Representation in Bicategories of Relations. ArXiv. 1706.00526v1 [cs.AI] 2 Jun 2017.

Paulheim, H. and F. Probst (2010). Application integration on the user interface level: An ontology-based approach. Data and Knowledge Engineering, 69, 1103-1116.

Penrose, Roger (1971). Applications of negative dimensional tensors. Combinatorial mathematics and its applications, 221244.

Penrose, R.; Rindler, W. (1984). Spinors and Space-Time: Vol I, Two-Spinor Calculus and Relativistic Fields. Cambridge University Press. pp. 424-425.

Sáenz-Ludlow, Adalira and Gert Kadunz (2015). Semiotics as a Tool for Learning Mathematics. Berlin: Springer.

Shin, S-J. (1994) The Logical Status of Diagrams, Cambridge: Cambridge University Press.

Sobociński, Pawel. Date accessed 15/11/19. Blog on Graphical Linear Algebra Blog. http://graphicallinearalgebra.net/.

Sowa, John F. (2000). Knowledge Representation: Logical, Philosophical, and Computational Foundations. Pacific Grove, CA: Brooks Cole Publishing.

Spivak, David I., Christina Vasilakopoulou,and Patrick Schultz (2019). Dynamical Systems and Sheaves. arXiv:1609.08086v4  [math.CT]  15 Mar 2019.Statebox Team, University of Oxford. Statebox. Date accessed 15/11/19. https://statebox.org/ .

Schultz, P., D. Spivak, C. Vasilakopoulou, & R. Wisnesky (2016). Algebraic Databases. arXiv:1602.03501v2 [math.CT] 15 Nov 2016.

Stjernfelt, Frederick (2007) Diagrammatology: An Investigation on the Borderlines of Phenomenology, Ontology, and Semiotics, Synthese Library, V. 336, Dordrecht, the Netherlands: Springer.

Vagner, D., Spivak, D. I. & E. Lerman (2014). Algebra of Open Systems on the Operad of Wiring Digrams, Date accessed 15/11/19. arXiv:1408.1598v1[math.CT] 7 Aug 2014.

Westrich, Q. (2006). Lie Algebras in Braided Monoidal Categories. Thesis, Karlstads Universitet, Karlstad, Sweden. http://www.diva-portal.org/smash/get/diva2:6050/FULLTEXT01.pdf

Willems, J.C. (2007). The behavioral approach to open and interconnected systems: Modeling by tearing, zooming, and linking. Control Systems Magazine, 27(46): 99.