Autonomy and Modularity



Autonomous systems of the kind that we can conceive as emerging from our technology are liable to be modular assemblages of elements that can couple opportunistically with other entities or systems, creating new assemblages whose powers and dispositions are transformed and dynamically put into play by such couplings.

The best way of representing modularity is in terms of networks consisting of nodes and their interconnections. A network is modular if it contains “highly interconnected clusters of nodes that are sparsely connected to nodes in other clusters” (Clunes, Mouret and Lipson 2013, 1). In autonomous assemblages modules support functional processes that make a distinct and specialized contribution to maintaining the conditions necessary for other interdependent processes within the assemblage.

Modules may or may not be spatially localized entities. They may be relatively fragmented while exhibiting dynamical cohesion. An instance of a software object class such as an “array” (an indexed list of objects of a single type) need not be instantiated on continuous regions on a computer’s physical memory. It does not matter where the data representing the array’s contents is stored is physically located so long as the more complex program which it composes can locate that data when it needs it. Thus while it is possible that all assemblages must have some spatially bounded parts – organelles in eukaryotic cells and distributors in internal combustion engines come in spatially bounded packages, for example – not all functionally discrete parts of assemblages need be spatially discrete in the way that organelles are. Cultural entities such as technologies or symbols may consist of repeatable or iterable patterns rather than things and may be conceived as repeatable particular events than objects (Roden 2004). Yet in systems – such as socio-technical networks – whose components cued to recognize and respond to patterns, such entities can exert real causal influence by being repeated into varying contexts.

Importantly for our purposes, dynamical cohesion should not be conflated with functional stability. An entity can retain its dynamical integrity and intrinsic powers while subtending distinct wide functional roles in the systems to which it belongs. To use, Don Ihde’s term: such entities are functionally “multistable”. An Acheulian hand axe – a technology used by humans for over a million years – might have been used as a scraper, a chopper or a projectile weapon.[1] Modern technologies such as mobile phones and computers are, of course, designed to be multistable; though their uses can exceed the specifications of their designers, as when a phone is used as a bomb detonator (Ihde 2012). It seems as if the decomposability of cognitive systems also confers multistability upon their parts thus contributing to the functional autonomy of the system as a whole.

In cognitive science, the classical modularity thesis held that human and animal minds contain encapsulated, fast and dirty, automatic (mandatory) domain-specific cognitive systems dedicated to specialized tasks such as kinship-evaluation, sentence-parsing or classifying life forms. However, it is an empirical question whether the mind is wholly or partly composed of domain-specific cognitive agents and, as Keith Frankish notes, a further empirical question whether neural modularity also holds: that is, whether domain-specified cognitive functions map onto anatomically discrete brain regions in the human brain such as Broca’s area (traditionally associated with language processing) or the so-called “Fusiform Face Area” (Frankish 2012, 280). Neither the classical theory of mental modules nor the neural modularity thesis follows from the fact that human brains are decomposable in the network sense presupposed by assemblage theory.

We should nonetheless expect autonomous entities such as present organisms or hypothetical posthumans to be network-decomposable assemblages rather than systems in which every part is equally coupled with every other part because modularity confers flexibility on known kinds of adaptive system.[2] For example, in biological populations modularity is recognized as one of the necessary conditions of evolvability “an organism’s capacity to generate heritable phenotypic variation.” (Kirschner and Gehart 1998, 8420). Some biologists argue that the transition from prokaryotic cells (whose DNA is not contained in a nucleus) to more complex eukaryotic cells (who have nucleated DNA as well as more specialized subsystems such as organelles) was accompanied by a decoupling of the processes of RNA transcription and subsequent translation into proteins. This may have allowed noncoding (intronic) RNA to assume regulatory roles necessary for producing more complex organisms because the separation of sites allows the intronic RNA to be spliced out of the messenger RNA where it might otherwise disrupt the production of proteins. If, as seems to be the case, regulatory portions of intronic DNA and RNA are necessary for the production of higher organisms, then this articulation in DNA expression may have allowed the ancestor populations of complex multi-cellular organisms to explore gene-regulation possibilities without disabling protein expression (Ruiz-Mirazo, Kepa and Moreno 2012, 39; Mattick 2004).

The benefits of articulation apply at higher levels of organization in living beings for reasons that may hold for autonomous “proto-ex-artefacts” poised for disconnection. Nervous systems need to be “dynamically decoupled” from the environment that they map and represent because perception, learning and memory rely on establishing specialized information channels and long term synaptic connections in the face of changing environmental stimulation. This entails a capacity “for cells to step back from the manifold of ambient stimulus and to be prepared to pick and choose which stimulus to make salient and thus in so doing a capacity to enjoy an unprecedented level of internal autonomy” (Moss 2006 932–934; Ruiz-Mirazo, Kepa and Moreno 2012, 44).[3]

Network decomposition of internal components also seems to carry advantages within control systems, including those that might actuate posthumans one day.  Research into locomotion in insects and arthropods shows that far from using a central control system to co-ordinate all the legs in a body, each leg tends to have its own pattern generator.



A coherent motion capable of supporting the body emerges from the excitatory and inhibitory actions of the distributed system rather than through co-ordination by a central controller. The evolutionary rationale for distributed control of locomotion can be painted in similar terms to that of the articulation of DNA transcription and expression considered above – a distributed system being far less fragile in the face of evolutionary tinkering than a central control architecture in which the function of each part is heavily dependent on those of other parts.

This rationale plausibly applies to human beings and as well as to our immediate primate ancestors, especially in the case of sophisticate cognitive feats that require the organism to learn specific cultural patterns – such as languages – which would not have been stable or invariant enough to have selected for the component abilities that they require over evolutionary time (Deacon 1997, 322-334  – the Visual Word Form Area is a particularly spectacular example of such “cultural recycling” – see below). While this is compatible with network decomposition it may not tally with the classical modularity thesis since it suggests an evolutionary rationale for the promiscuous re-use of functionally multistable components.

Evidence from functional imaging suggests that anatomically discrete regions like Broca’s or the Fusiform Area are co-opted by evolutionary and cultural processes in support of functionally disparate cognitive tasks. For example, relatively ancient areas in the human brain known to be involved in motor control are also involved in language understanding. This suggests that circuits associated with grasping the affordances and potentialities of objects were recruited over evolutionary time to meet the emerging cultural demands of symbolic communication (Anderson 2007, 14). In a recent target article on neural-reuse in Behavioural and Brain Sciences Michael Anderson cites research suggesting that older brain areas tend to be less domain specific and more multistable – that is, that they tend to get re-deployed in a wider variety of cognitive domains (Anderson 2010, 247). Peter Carruthers and Keith Frankish likewise argue that circuits in the visual and motor areas which have been initially involved in controlling and anticipating actions have become co-opted in the production and monitoring of propositional thinking (beliefs, desires, intentions, etc.) through the production of inner speech. A an explicit belief, for example, can be implemented as a globally available action-representation – an offline “rehearsal” of a verbal utterance – to which distinctive commitments to further action or inference can be undertaken (Carruthers 2008). Andy Clark cites experimental work on Pan troglodytes chimpanzees which comports with the Carruthers’ and Frankish’s assumption that cognitive systems adapted for pattern recognition and motor control can be opportunistically reused to bootstrap an organism’s cognitive abilities. Here, an experimental group of chimps were trained to associate two different plastic tokens with pairs of identical and pairs of different objects respectively. The experimental group were later able to solve a difficult second-order difference categorization task that defeated the control group of chimps who had not been trained to use the tokens:

The more abstract problem (which even we sometimes find initially difficult!) is to categorize pairs-of pairs of objects in terms of higher order sameness or different. Thus the appropriate judgement for pair-of-pairs “shoe/shoe and banana/shoe” is “different” because the relations exhibited within each pair are different. In shoe/shoe the (lower order) relation is “sameness”; in banana/shoe it is difference. Hence the higher-order relation – the relation between the relations – is difference (Clark 2003, 70).

Interestingly, Clark notes that the chimps in the experimental group were able to solve the problem without repeatedly using the physical tokens, suggesting that they were able to associate the “difference” and “sameness” with inner surrogates similar to the offline speech events posited by Carruthers and Frankish (71; See also Wheeler 2004).

This account of the emergence of specialized symbolic thinking and linguistic thinking via the reuse of neural circuits evolved for pattern recognition and motor-control illustrates a more general ontological schema. Assemblages – whether human, inhuman, animate or inanimate – inherit the capacity to couple with larger assemblages from their structure and components and are similarly constrained by those powers. Carbon atoms have the power to assemble complex molecular chains because their four valence electrons permit the formation of multiple chemical bonds. Simpler prokaryotic cells may lack the capacity to evolve the regulatory networks required to form multicellular affiliations because their encoding process is insufficiently differentiated. Likewise, although specific neural circuits may be inherently multistable it does not follow that each can do anything. Each may have specific “biases” or computational powers that reflect its evolutionary origins (Anderson 2010, 247). For example, Stanislas Dehaene and Laurent Cohen review some remarkable results suggesting the existence of a Visual Word Form Area, a culturally universal cortical map situated in the fusiform gyrus of temporal lobe, which is involved in the recognition of discrete and complex written characters independently of writing system.

As Dehaene and Cohen observe, it is not plausible to suppose that the VWFA evolved specifically to meet the demands of literate cultures since writing was invented only 5400 years ago, while only a fraction of humans have been able to read for most of this period (Dehaene and Cohen 2007, 384). Thus it appears that the cortical maps in the VWFA have structural properties which make them ideal for reuse in script recognition despite not having evolved for the representation of written characters (among the factors suggested is that the VWFA is located in a part of the Fusiform area receptive to highly discriminate visual input from the fovea – 389).

Coupling an assemblage with another system – e.g. a transcultural code such as a writing or number system – may, of course, increase the functional autonomy of system by allowing it to respond fluidly and adaptively to the demands of its environment – enlisting new affiliations and resources which, then, come to be functional for it. Literacy and numeracy have become functionally necessary for economic activity in advanced industrial societies – clearly this was not always so! However, this is only possible because both the assemblage and its parts are open to functional shifts that, in effect, allow the creation of new social “megamachines” which extend beyond the coupled individuals. Thus while complex assemblages articulated into lots of functionally open systems may be more functionally autonomous than less articulated ones – are more capable of accruing new functions –they are more apt to be “deterritorialized” by happening on new modes of existence and new ways of being affected (DeLanda 2006, 50-51).


Anderson, Michael (2007). “Massive redeployment, exaptation, and the functional integration of cognitive operations”. Synthese, 159(3): 329-345,

Anderson, M. L. (2010). “Neural reuse: A fundamental organizational principle of the brain.” Behavioral and Brain Sciences, 33(4), 245.

Carruthers, Peter (2008). “An architecture for dual reasoning”. In J. Evans & K. Frankish (eds.), In Two Minds: Dual Processes and Beyond. Oxford University Press.

Clark, Andy (2003). Natural Born Cyborgs. Oxford: Oxford University Press.

Clune, J., Mouret, J. B., & Lipson, H. (2012). “The evolutionary origins of modularity”. arXiv preprint arXiv:1207.2743.

Deacon, Terrence. 1997. The Symbolic Species: The Co-evolution of Language and the Human Brain . London: Penguin.

Dehaene, S., & Cohen, L. (2007). Cultural recycling of cortical maps. Neuron,56(2), 384-398.

DeLanda, M. (2006), A New Philosophy of Society: Assemblage Theory and Social Complexity, London: Continuum.

Frankish, Keith (2012). “Cognitive Capacities, Mental Modules, and Neural Regions”. Philosophy, Psychiatry, and Psychology 18 (4).

Ihde, D. (2012). “Can Continental Philosophy Deal with the New Technologies?” Journal Of Speculative Philosophy, 26(2), 321-332.

Kirschner Marc and Gehart, John (1998). “Evolvability”, Proceedings of the National Academy of Sciences USA, 95, 8420-8427.

Moss, L. (2006). “Redundancy, plasticity, and detachment: The implications of comparative genomics for evolutionary thinking”. Philosophy of Science, 73, 930–946.

Roden, David (2004). ‘Radical Quotation and Real Repetition’, Ratio (new series) XVII 2 June 2004, 191-206.

Ruiz-Mirazo, Kepa & Moreno, Alvaro (2012). “Autonomy in evolution: from minimal to complex life”. Synthese 185 (1):21-52.

Wheeler, M. (2004). “Is language the ultimate artefact?.” Language Sciences, 26(6), 693-715.


[1] See Don Ihde, “Embodiment and Multistability”,, accessed 14/02/2013.

[2] One of the benefits of so-called “objected oriented” programming languages (OO) like Java over “procedural” programming languages such as COBOL is that OO programs organize software objects in encapsulated modules. When a client object in the program has to access an object (e.g. a data structure such a list) it sends a message to the object that activates one of the objects “public” methods (e.g. the client might “tell” the object to return an element stored in it, add a new element or carry out an operation on existing elements). However, the client’s message does not specify how the operation is to be performed. This is specified in the code for the object. From the perspective of the client, the object is a black box that can be activated by public messages yielding a consumable output. This means that changes in how the proprietary methods of the object are implemented do not force developers to change the code in other parts of the program since these do not “matter” to the other objects. Maintenance and development of software systems becomes simpler.

[3] The cochlear cells in our inner are connected to hair like cells which are receptive to sound vibrations. This specialized arrangement allows the cochlear to conduct a fast spectrum analysis on incoming vibrations, assaying the relative amplitudes of components in complex sounds.

Leave a Reply

Fill in your details below or click an icon to log in: Logo

You are commenting using your account. Log Out /  Change )

Twitter picture

You are commenting using your Twitter account. Log Out /  Change )

Facebook photo

You are commenting using your Facebook account. Log Out /  Change )

Connecting to %s