Download Enriching Attachment Models: From Internal to Embodied Working Models and more Summaries Psychoanalysis in PDF only on Docsity!
From Internal Working Models to Embodied Working
Models
Dean Petters 1 and Everett Waters 2
Abstract. John Bowlby introduced the ‘Internal Working Model’ construct into Attachment Theory to explain attachment phenomena such as an individual making plans and predictions about how attachment set-goals can be achieved. For example, when an attached individual in an anxious state considers different ways to gain proximity (and hence security) to their attachment figure before acting. According to Bowlby, an individual can possess multiple Internal Working Mod- els, which can differ in many respects, such as representational for- mat or how much the individual is aware of them. Existing agent based attachment simulations have implemented Internal Working Models but these are greatly simplified in comparison with Bowlby’s rich and diverse conceptualisation. This paper proposes that com- putational attachment models can be enriched by: (i) incorporating and adapting the idea from the psychoanalytic tradition that experi- ence can be recorded in terms of how structures of mental energy and defence are built rather than representing key attachment experi- ences in memory. So the particular architectural formations that de- velop are the residue of experience that can bias and filter future pro- cessing; and (ii) viewing cognitive architectures as more contingent and ephemeral in their structure than typically conceived. So that ar- chitectures are viewed as not just switching between configurations quickly, like the attractor states in a dynamical system, but doing so in a way that captures and brings to bear coherent biases and filters in processing laid down over long term development.
1 Introduction
John Bowlby’s interest in developmental psychology started early in his career [10, 23]. After working with maladjusted children, he was training as a medical doctor when he added psychoanalysis to his studies. Melanie Klein acted as his supervisor during this psychoan- alytic training. Bowlby went onto develop Attachment Theory as a theoretical vehicle to conserve some of the key insights of psycho- analysis whilst abandoning some of the aspects of the psychoanalytic framework with which he disagreed. The aspects of psychoanalytic explanation that he wished to conserve included that the cognitive and emotional life of human infants is complex and that the nature of early attachment relationships have a lasting impact, acting as proto- types of later romantic and caregiving relationships [25]. However, Bowlby disagreed with the mental energy and drive reduction models that psychoanalysis proposed to explain such internal complexity and continuity across development [25]. Other elements of psychoana- lytic explanation that Bowlby wished to conserve for developmental psychology included that the phenomena of interest are bigger than
(^1) University of Northampton, UK, email: [email protected] (^2) SUNY, Stony Brook, USA.
the ‘proxy’ of behaviour. For both Psychoanalysis and Attachment Theory, overt behaviour (for example, duration of protest following separation) does not equate with strength of emotional connection [25]. In both of these frameworks, responses are guided by rich in- ternal structures and mechanisms. However, Bowlby placed far more emphasis on the observation of current behaviour than did Melanie Klein and other psychoanalysts, who emphasised the retrospective research method of clinical reconstructions. So Bowlby viewed the immediate context an individual is in, who is around them, and their immediately prior experiences, as more important influences on their behavioural responses. The tension between the conflicting views of how research should be conducted and the relative importance of ob- serving actual behaviour is illustrated in this passage from van der Horst:
“[Bowlby] was seeing an anxious, hyperactive child as a patient five days a week. The boy’s mother would sit in the waiting room, and Bowlby noticed that she too seemed quite anxious and unhappy. When he told Klein he wanted to talk to the mother as well, Klein refused adamantly, dismissing the mother as a possible causal or related factor in the child’s be- haviour, and saying “Dr Bowlby, we are not concerned with reality, we are concerned only with the fantasy” (Kagan, 2006, p 43). When the mother was subsequently taken to a mental hospital for treatment of anxiety and depression, Klein was un- affected and untouched and only replied that is was a nuisance because now they had “to find another case” (Karen 1994, p 46). Bowlby was thoroughly annoyed - even 50 years later, in a conversation with the well-known developmental psychologist Jerome Kagan, he still become angry when relating this case (Kagan 2006) and distanced himself from Klein.”. Many years later Bowlby describe his own view that:
“most of what goes on in the internal world is a more or less accurate reflection of what an individual has experienced recently or long ago in the external world. Of course, in addition to all that, we imagine things ... but most of the time we’re concerned with ordinary events. If a child sees his mother as a very loving person, the changes are that his mother is a loving person. If he sees her as a rejecting person, she is a very rejecting person” (Bowlby, Figlio and Young, 1986, p. 43) ([23], p 21 )
So Bowlby was principally focused on how we represent our day to day experiences and use those to make predictions about future outcomes. Bowlby did want to consider how individuals imagine good and bad future outcomes. He just believed that the reality of the current moment and real past experiences around emotionally
valenced possible attachment outcomes like loss, separation and re- union anchor imaginative ’what if’ reasoning when an individual looks ahead to possible futures. However, whilst keeping in mind attachment behaviour was a stand-in (shortcut) for the richer inter- nal and hidden phenomena he wanted to study, Bowlby also some- times treated behaviour as the phenomenon to be studied. Focusing on the study of behaviour itself (within an ethological framework) was methodologically valuable because it focused research on the part of the problem that is empirically accessible. It was also strate- gically valuable because it allowed Bowlby to break with psychoan- alytic clinical reconstructions as a research methodology. So in part, Bowlby emphasised the importance of observation of actual current behaviour for pragmatic reasons, not because there was not more to consider.
2 Internal Working Models as a scientifically
respectable concept
Bowlby’s recognition of the tension between the need for both con- servation and change in regard to psychoanalytic constructs led him to propose an alternative motivational model based on the etholog- ical and control systems theories of the day ([7], chapter 1). At the centre of the attachment control system framework is the Internal Working Model (IWM), a more scientifically respectable construct than psychoanalytic drives and psychic energy. IWMs, in their pro- posed structure and operation, were compatible with the emerging information processing frameworks in Cybernetics, Artificial Intelli- gence and Cognitive Psychology. The conceptual ancestry of IWMs goes back indirectly to Craik’s (1943) proposal of ‘Working Mod- els’ [8]. In the broader sense in which Craik used the term, Working Models are not confined to attachment but apply to all representative models of the world. In his work Bowlby restricted the term Internal Working Models (IWMs) to models of self and other in attachment relationships. IWMs capture the relation-structure of attachment phe- nomena, not every aspect of reality but enough to make possible the evaluation of alternative actions. These include spatio-temporal causal relations among the events, actions, objects, goals and con- cepts represented. IWMs of attachment are what hold an infant’s ex- pectations of the levels of predicted availability and responsiveness for a given carer. These expectations are derived from the carers past performance. IWMs of self and attachment figure develop in a com- plementary manner. For example if the carer is responsive the self is valued. Their operation can be seen when an attached individual is in an anxious state and considers how to gain proximity to their at- tachment figure. IWMs allow the individual to predict the outcomes of possible actions to achieve their set-goal of proximity. They can then choose an action likely to increase security and not provoke a negative response from their attachment figure. Although Bowlby used the IWM concept more narrowly than Craik in confining it more to just the attachment context, he also framed the IWM more broadly than Craik’s working models. For Craik, the working models which a living organism might possess in their minds were comparable to the physical systems which scientists used to explain natural phenomena:
“By a model we thus mean any physical or chemical system which has a similar relation-structure to that of the processes it imitates. By ’relation-structure’ I do not mean some obscure non-physical entity which attends the model, but the fact that it is a physical working model which works in the same way as the process it parallels, in the aspects under consideration at
any moment. Thus, the model need not resemble the real object pictorially; Kelvin’s tide-predictor, which consists of a number of pulleys on levers, does not resemble a tide in appearance, but it works in the same way in certain essential respects”([8], p 51)
Craik’s working models are physical systems which can act as models to explain natural phenomena because their physical oper- ation captures key aspects of how the target system operates. When an organism holds a working model in its mind which represent its self and environment, it can configure the working model to act as memories of past events and then run this model forward in time to make predictions or imagine the results of differing actions:
“If the organisms carries a ‘small scale model’ of exter- nal reality and if its own possible actions within its head, it is able to try out various alternatives, conclude which is the best of them, react to future situations before they arise, utilise the knowledge of past events in dealing with the present and future, and in every way react in a much fuller, safer, and more com- petent manner to the emergencies which face it” ([8], p 61).
Bowlby himself sometimes presented IWMs in this way, for ex- ample suggesting that Internal Working Models allow “small scale experiments” to be conducted “within the head” ([7], p 80). How- ever, it would be a mistake to go back to Craik’s ideas about work- ing models and interpret IWMs as experienced like a simulation of a visual scene, rather than the more abstract variant Craik proposed where it is the similar ‘relation structure’ that is key. This narrow ‘visual-image’ concept for what IWMs might be would be quite slow for adults for use in dynamic situations, and would not be available to sensorimotor infants. Although many attachment theorists think of IWMs as ‘the’ attachment representation, a more developed treat- ment is to instead view IWMs as referring to a category of internal representations. It is worth emphasizing that in the examples above which Craik gives for working models, although these systems can be argued to symbolize reality, it is by their physical properties rather than with abstract or arbitrary symbols that they represent other systems. This can be contrasted with Bowlby who proposed IWMs to include ana- logue and sensorimotor variants, but also to be represented as inter- nal symbols and even interface with other high level processes of integration and control like natural language ([7], pp 81-82). This is an important distinction. Attachment Theory has moved away from more narrow conceptualisations of Working Models and proposes that there are multiple varieties with differing modes of representa- tion. So a key issue for computational modelling is whether IWMs are a single type of representation or a diverse class of representa- tions that afford prospective capabilities. If a diverse class if IWMs is required, what range of designs will these possess?
3 Internal Working Models versus Inner worlds:
imagined futures and subjective experience
Bowlby compared IWMs to the Internal Worlds of psychoanalysis:
“The environmental and organismic models described here as necessary parts of a sophisticated biological control system are, of course, none other than the internal worlds of traditional psychoanalytic theory seen in a new perspective.” ([7], p 81)
However, Holmes, as psychoanalytically inclined Attachment Theorists observes:
tor social-interactions in infancy, (like sinking in when held), to Inter- nal Working Models in later childhood and adulthood which may be symbolically or linguistically mediated. A simple view is that an in- dividual progresses from relying on embodied attachment represen- tations early in development to higher level representations in later development. However, even in adulthood very different levels of at- tachment representation may interact [20]. Computational attachment models implemented as robotic and agent based software simulations show a range of possibilities, from cybernetic inspired control systems [6] and behaviour based archi- tectures [14, 19, 2, 9] to implementing Internal Working Models in agent based simulations as internal subsystems which operate over symbolic representations [19, 18]. The symbolic systems in [19, 18] (illustrated in figure 1 and implemented in pop11 using the sim-agent toolkit) deliberate about the likely future outcome of actions by sys- tematically searching and appraising possible actions and outcomes ([18], pp 103-152).
Figure 1. A hybrid attachment architecture with reactive and deliberative subsystems that has been implemented in pop11 using the sim-agent toolkit [19, 18]. The lower reactive behaviour system has a winner take all action selection system which decides which movement and signalling actions to take. In parallel, the deliberative and meta-management component in the architecture uses the same perceptual data and goals to create, evaluate, select and execute actions, which include inhibition of the actions output by the reactive behaviour based subsystem (inhibiting actions denoted with round-headed information flow). So this deliberative and meta-management component can be viewed as an Internal Working Model that ‘allows small scale experiments to be conducted within the head’ before external actions are taken.
The symbolic subsystem in the deliberative and meta-management components of the architecture in figure 1 operates in parallel with a behaviour based subsystem which receives inputs from goal activa- tors for safety, exploration, socialisation and physical need. In the be- haviour based subsystem the goal with the highest activation directs the next movement and signalling actions. What the symbolic IWM does in this architecture is receive the same inputs and then produce plans towards the same goals as the lower behaviour based system. When the best plans from the IWM suggest actions that conflict with the actions activated by the lower behaviour system route, the upper system can intervene and inhibit the operation of the lower system. So if the lower system activates moving and signalling the upper sys- tem may inhibit the signalling and just leave the activated movement. There are many aspects of Bowlby’s conceptualisation for IWMs this
simulation cannot model - it only possesses one IWM rather than the multiple varieties proposed by Bowlby; it can only interact with its environment through perception and action at a reactive level (so not perceptions or actions of language or with ‘short-hand’ references to mental states or other ‘hidden’ aspects of the environment; it can only reason from goals selected by the lower behaviour based sys- tem, rather than reason about what are good goals to hold; and of most relevance for the ongoing argument in this paper, this attach- ment simulation does not attempt the explain the subjective experi- ences which are linked with the behaviour or cognitive elements it models. So what is missing from Bowlby’s account and contemporary computational models is some clear explicit mechanistic causal framework that links all these three elements of behaviour, cogni- tion, and subjective experience together. It is clear that links exist between behaviour, cognition, and subjective experience. However, exactly what these links are is not clear. This issue is not isolated to Attachment Theory but is an example of a more common problem which has been termed: “the explanatory gap” between sub-personal computational cognition and subjective mental phenomena ([22], p 6).
5 Embodied Working Models
Progress towards solving the problems of modelling subjective expe- rience and in producing systems with more diverse ways to internally model the self, environment and attachment may be made by taking an embodied view of attachment phenomena. In this approach the body can be modelled as the context or milieu of attachment struc- tures and mechanisms. The cognitive component of Attachment The- ory could then be augmented with the incorporation of bodily sensa- tions, physiological responses, and analogue computations that rely on the physical substrate within the attachment control system. So an embodied approach might then encompass the body as a lived expe- riential structure ( [24], p xvi) resolving questions raised by the cur- rent use of overly cognitive and internal attachment representations in computational attachment modelling. In this view, the subjective feelings associated with attachment episodes and the prospections that occur about episodes ‘yet to be’ could be conceptualised as be- ing brought forth from a history of structural coupling ([24] p 205).
6 Recording experience as structure not encoded
memory
This section is concerned with several related questions: how are early experiences recorded in a way that they impact later experi- ences?; what changes as a result of previous experience?; and what changes during different timescales, over days, weeks or months? Ainsworth explored this issue when she discussed how what occurs over previous days may affect the sub-categories found in the Strange Situation [1]. One answer is the infants persistent state. When a baby has become anxious or worried, ‘on edge’, this might be measured with a cortisol assessment. This is not the representation of informa- tion in memory. A second answer is that the baby has attached some kind of meaning to the experience. This ‘meaning making’ reaction might be mediated through information processing mechanisms such as encoding into memory that works through a longer period of time than physiological state. There is also another longer term way of recording experience which is that there is structural change in the form of the whole information processing architecture. This latter
suggestion for how experience is recorded comes from the psychoan- alytic tradition and is the idea that experience was recorded in terms of building structures of mental energy and defence ([11] p 17). Lit- tle dykes to keep drives from flooding off in certain directions and ditches o direct energy in other directions. So structure is the residue of experience that filters and biases future behaviour. One of Bowlby’s contributions in the 1960s in his revision of the psychoanalytic framework was to introduce the idea of information, rather than structure, as the means of recording past experience. The notion of bolting a memory or simulation module like an IWM on only really works with these information processing formulations for the attachment control system. In this ‘Back to the Future’ view proposed here (that draws upon old psychoanalytic ideas in a new context), the attachment control system grows in a way that records what it experiences during its de- velopment. The elements (building blocks) for an attachment control system, as a result of experience, become coordinated into a system to give a particular structural form. In the same way that a tree planted on a hill with a strong directed wind will grow leaning and so record the prevalent wind direction. In this new view, the word architecture is a little misleading as a label because it suggests a structure (like a building) that is designed towards a blue-print, or some design in the mind of the architect. Instead we might think about landscapes with a 3D structure that changes over time. So like a building com- pleted over many generations. Or perhaps the analogy of experience being like water running over a landscape and eroding gulleys nicely captures psychoanalytic view of how structure can gets formed by existing within its environment. Or another analogy is a cognitive ar- chitecture like a living, flexible organ like a heart, with blood being pumped around. In this analogy, experience is like triggers for valves opening and closing and blood pushed one way or the other and some routes becoming preferred/more likely than others. What does this mean for the imagining of possible futures? Per- haps with particular experiences an architecture takes a form that prospects just ’pop-out’ in a way that does not require fully deliber- ative mechanisms. So the future that comes to be imagined does so because of architectural structure as much as the contents of mem- ory. Such a component would be situated at the deliberative level but in the perceptual column of figure 1. The same may occur with mental state ascription - the architectural structure as well as the con- tents of memory promoting particular interpretations of what others are thinking and feeling. Mind-reading due to the right architectural formations rather than the right algorithmic operations on the con- tents of memory. Such a component would be situated at the meta- management level but in the perceptual column of figure 1. This ‘ar- chitecture based imagination’ concept has similiarities with Sloman’s idea of ‘architecture based motivation’ [21], where the architectural organisation rather than some response to reinforcement or reward signals can decides motives. How does this view relate to the scientific objectives of cognitive modelling? When an attachment modeller wants to implement a sys- tem that encodes experience in memory they are on familiar territory as computers naturally use memory encoding. Having an architec- ture that changes in the way that a landscape or living organ does is much further from traditional digital computing. So there is a bias in people who do modelling. However, we should not have theories of development get produced because they seem manageable from an AI techniques perspective. What is required is an unbiased view of the modelling outcome and then try and fit to what can currently be modelled, aware of any possible shortfall. Otherwise what will be produced is a mere simulation rather than a very deep one.
7 Viewing architectures as ephemeral structures
with affordances that change moment to moment
Newell (1990) defined a cognitive architecture as:
“the fixed (or slowly varying) structure that forms the framework for the immediate processes of cognitive perfor- mance and learning.” (Newell 1990, p 12)
This section will present an argument that a more fine grained view of architectural temporality has some benefits when we move from a model that has a central processor and passive memory stores to a model where experience and biases are recorded in architectural structure. To unwrap this issue we will first consider a high level cog- nitive architecture (of a generic type similar to ACT-R [4], or EPIC [15, 16]) which processes information with productions rules. Some production rules will be involved in deliberative processes and oth- ers in meta-management processes. In dual task activities the meta- management processes may be very busy, working out when to swap from one deliberative task to another. However, in terms of resource constraints, both the meta process and the deliberative process both use attention requiring resources, they are both resource constrained. So the architecture cannot deliberate and meta-process at the same time. How does the architecture ‘decide’ which process to run next? At an implementational level, this means deciding which production rule to fire next when only one can fire. Older versions of cognitive architectures based upon production rules might have a goal stack to mandate the order in which goals and sub-goals should be pro- cessed [3, 5]. With a goal stack, everything that would happen would be predictable. The control structures would just go down the goal stack and deal with whatever process needs to be carried out as its turn arrive. However, goal stacks are not very biologically plausible and are absent in contemporary production system architectures that aspire for biological plausibility (such as ACT-R 6, [4]). Instead, the ‘correct’ production in some sequence is primed or activated rather than strongly directed. The goals are dynamically constructed. This is ‘soft power’ rather than ‘hard power’. However, because of this processes can get swapped about. From the perspective of one sin- gle expected/desired chain of processes in a particular task, if all the particular parts were not quite right the ‘wrong’ thing may happen. Priming and spreading activation help form the context to trigger the next correct action. What priming and spreading activation do is change what the architecture is - in terms of not only what states are active, but in terms of what states are accessible from the current states. They change what the architecture affords, at that moment, in terms of it possibilities. What this means is that if we are coming up with metaphors for what a cognitive architecture is, we have think about more dynamic metaphors than an architecture which is similar to a building like a castle. Cognitive architectures are more contingent on the particu- lar context. If a cognitive architectures is compared to building with rooms, then it is a strange building where rooms that rise or lower so you cannot get in some rooms until they have appeared (perhaps by fast-acting hydraulics!). So there is an ephemeral nature to the organ- isation of a cognitive architecture - perhaps better termed ‘ephemeral structures’. Instead of thinking about a cognitive architecture like a castle we should perhaps think if it like a special children’s bouncy castle. This gets blown up and collapses and blown up again slightly differently. Perhaps the turrets have all moved around affording a completely different set of processes to possibly become active. In the context of ‘what if’ reasoning, and explaining the imagined outcomes that can arise from this, it is not just your upbringing that
ogy’, AISB Convention 2010, 51–58, AISB Press, University of Sussex, Brighton, (2010). [21] A. Sloman, ‘Architecture-Based Motivation vs Reward-Based Motiva- tion’, Newsletter on Philosophy and Computers, 09 (1), 10–13, (2009). [22] E. Thompson, Mind in Life: Biology, Phenomenology, and the Sciences of Mind, MIT Press,, Cambridge, Mass, 2007. [23] F. van der Horst, John Bowlby - From Psychoanalysis to Ethology: Unravelling the Roots of Attachment Theory, Wiley-Blackwell, Chich- ester, 2011. [24] F. Varela, E. Thompson, and E. Rosch, The Embodied Mind: Cognitive Science and Human Experience., MIT Press,, Cambridge, Mass, 1991. [25] E. Waters, K. Kondo-Ikemura, G. Posada, and J. Richters, ‘Learning to love: Mechanisms and milestones’, in Minnesota Symposium on Child Psychology (Vol. 23: Self Processes and Development), eds. M. Gunner & Alan Sroufe, 217–255, Psychology Press, Florence, KY, (1991).
