The Fables of Lucy R.:
Association and Dissociation in Neural Networks
Department of Philosophy
Program in Neuroscience
Hartford, CT 06106
According to Aristotle, "to be learning something is the greatest of pleasures not only to the philosopher but also to the rest of mankind," (Poetics 1448b). But even as he affirms the unbounded human capacity for integrating new experience with existing knowledge, he alludes to a significant exception: "The sight of certain things gives us pain, but we enjoy looking at the most exact images of them, whether the forms of animals which we greatly despise or of corpses." Our capacity for learning is happily engaged in viewing representations of painful objects, but not, it seems, in viewing the objects themselves. When an experience is intensely painful, what then is a rational animal to do? We can neither disable our learning process, nor erase its traces. In the face of intense pain, horror, or terror, learning and remembrance cause no pleasure but rather persistent psychological pain and disruption. The memorious mind reverberates with trauma.
The traumatized mind responds in diverse ways to the recurrent crises of reminiscence, responses which lead at the extreme to the symptoms of various disorders. These reactions fall into two broad camps, the associative and the dissociative. The first is exemplified by some (but not all) of the symptoms of post-traumatic stress disorder, in cases where even a trivial element associated with the painful event becomes an evocative cue for reliving the experience. In contrast, dissociation is characterized by subjective distancing from the initial pain and its remembrance, often with secondary effects. In dissociative amnesia, for example, subjects fail to recall critical spans of their lives, often seeming to obliterate the traumatic memory. The erasure is only apparent; in diverse ways the trauma continues to oppress even if it cannot be consciously recalled.
There are, or course, many (too many) occasions of trauma. That diversity, and the diversity of responses in its aftermath, imply that the causal mechanisms of traumatized cognition are manifold. Understanding post-traumatic psychopathology is further complicated by the compounded effects of multiple or repeated trauma. With this complexity in mind, in this chapter I will explore connectionism as a unifying framework for understanding the traumatized mind. The first motive for this attempt is already apparent. Trauma is an occasion for a kind of learning, and connectionist models are most adept at simulating learning. In addition, a connectionist model offers extraordinary flexibility in representation. Arrays of neural processing units afford subtlety and precision in simulating the contents of mind. With learning, these representations change. "Traumatic learning" can thus be modelled, and a network observed in its initial responses, and then subjected to further simulated trauma with further testing. In this way, a narrative of traumatic experience and its diverse psychological manifestations can be condensed, simplified, and examined. Even a simple network allows many variations. For a first foray into the simulation of psychogenic psychopathology, I followed a well-known case study, using the concrete history and experienced symptoms of a patient (and her therapist) as a guide for a network model. The case study is exemplary of the 90s -- the 1890s, that is. In various ways it is emblematic of a century of clinical thinking that followed.
Sometime in the fall of 1892 a governess working in the outskirts of Vienna visited her doctor with an unusual complex of symptoms. The patient presented a physical symptom, a chronic suppurative rhinitis, combined with a "psychological" symptom, a persistent olfactory hallucination, the smell of burnt pudding. Her doctor referred the case to Sigmund Freud, who ultimately told the patient's story as the case study of "Lucy R.," in Freud and Breuer's Studies on Hysteria (1955 (1895)). There Freud described the case as "a model instance of one particular type of hysteria, namely the form of this illness which can be acquired even by a person of sound heredity, as a result of appropriate experiences" (1955 (1895): 122). Looking back on the case from the other end of Freud's career, it seems to be a model instance of more than just a type of hysteria. It anticipates Freud's own evolution toward psychoanalysis, as well as developments in clinical approaches to psychopathology both before and after Freud. In the early 90's, Freud and Breuer hypothesized that "hysterics suffer mainly from reminiscences."(7) In the Studies, the mind of the hysteric is seen to be fending off painful memories and wishes, with unintended symptomatic side-effects. This basic process of thoughts repressed and resurgent would soon be applied in pathologies other than hysteria, and ultimately be read into every detail of 20th century life. Even more ubiquitous in our time is the presupposition implied by repression, namely, that there is something to repress, in the form of thoughts exerting causal powers without entering consciousness. The conception of an active-but-unconscious realm of mind reigns in both clinical and cognitive psychology. The study of Lucy R. explicates these ideas in terms that are familiar still. In addition, the psychoanalytic method seems already implicitly at work with Lucy R. Unlike Breuer's treatment of Anna O., for example, with Lucy Freud abandoned hypnotism as a diagnostic and therapeutic aid, probing instead for remembered associations and achieving a "talking cure."
Yet at the same time the case of Lucy R. is not yet laden with apparatus and theory that Freud would later develop. Ego, Id, Superego, and the specific complexes from infancy all lay in Freud's future. More important, in the Studies Freud still clearly conceived of psychopathology as originating in traumatic experience. Later, he would relocate the origins of pathology in repressed fantasies and wishes, a reorientation that decisively influenced the subsequent development of psychoanalysis. In the earlier, "experiential," conception, Freud showed his affinity with the great 19th century theorists Charcot (with whom Freud studied) and Janet. And, conveniently, this orientation has reemerged as a central contemporary issue. Although hysteria has fallen out of terminological favor, both its symptoms and traumatic etiology still echo in contemporary clinical taxonomy and theory (Kihlstrom 1994). For the dissociative disorders in particular, a frequent cause is real, not merely imagined, trauma. Lucy R., as Freud interpreted her experience, provides a straightforward example.
For these reasons I turned to Lucy R. as a model to approach within the connectionist framework. As in all of his case studies, Freud used his subject as an object lesson from which he drew specific conclusions. As we revisit Freud's fable, we will examine the conclusions he drew and draw some new ones as well.
simulation of Lucy R.
Representing Lucy R.'s phenomenology
Lucy R. also enjoys the curious distinction of having been the target of an earlier attempted computer simulation, which was described but not implemented by Cornelis Wegman (1985). In the epilogue to his heroic effort, Wegman estimated that the actual implementation of the model would require at least five years of programming work. How did the case of Lucy R., one of Freud's briefest, get so complicated? The complexity, I suggest, reflects the limitations the approach taken, which was drawn entirely from classical Artificial Intelligence in its heyday, before the reemergence of connectionism. A brief look at Wegman's work shows the limits of classical AI as a concrete modelling tool, suggests the unique powers of connectionism, and offers an initial foot-hold on "Lucynet," our sequel to Freud's study.
Wegman based his efforts on a sophisticated theory of "knowledge structures" and their modifications, Roger Shank's theory of scripts (Shank and Abelson 1977). A script is a framework for knowledge about actions or other sequences of events. To define a script, one draws from a repertoire of actions (of which some are basic), each action having specified effects on the world. The problem unfolds from the words "define" and "specify": Every change, and all its relevant or probable consequences, must be programmed by hand. Any genuine interaction in the world demands a huge script for its description. The one event from the Lucy saga that Wegman fully scripts accordingly implicates seventeen separate actions fulfilling ten distinct goals and plans for their achievement. Being an outline and not an implemented program, there is no guarantee that even this level of detail will be enough. His initial scripting is certainly not enough, Wegman notes, to incorporate the proposed mechanisms of hysteria, and thence he carefully and explicitly adds each of the following to the expanding model: Affect, arousal, facial expression, abreaction, working-over, episodic memory, and several others. Significantly, a separate component called "consciousness" appears in the flow chart in order to enable the functions of attention and repression. (For a similar exegesis, but much less sympathetic to Freud, see Cummins 1983.)
The exercise has the enormous value of flushing out unnoticed ambiguities is Freud's hypotheses. But even if the model led to a working implementation, it would leave open the question which if any of the many installed modules was real. Like the epicycles of Ptolemaic astronomy, every module in this sort of model is a kludge, hand-built to do just what Freud (as interpreted by the would-be modeler) supposes. A more powerful demonstration, in contrast, would be one that requires less programming intervention, operates according to a few highly general principles, and yet exhibits several of the symptoms and responses in parallel with the case study. A model of this sort has the capacity to exhibit the target phenomena as emergent side-effects of more basic and general functions, rather than as the explicit result of a deliberate program. This in turn suggests new hypotheses and tests in the target domain. Connectionist models, when they work, have these attractive features. Indeed, connectionism makes a virtue of simplicity. The simpler the model, the more basic and general will be the mechanisms explaining its functional behavior, resulting in more powerful explanatory hypotheses in the real world.
To construct Lucynet, then, I took the minimalist approach of radical simplicity, as afforded by connectionism. Network design was guided by the observations reported by Freud, which I took at face value (pacé Crews 1993, Grünbaum 1984, and other critics of Freud's self-fulfilling observations). Lucy's primary symptoms, unlike those reported elsewhere in the Studies, were perceptual hallucinations, a smell of burnt pudding and the smell of cigar smoke. To understand these symptoms, Freud probed Lucy's conscious memories and eventually uncovered episodes that Lucy had apparently repressed. But although Lucy would not easily recall them, these episodes were initially fully conscious. The goal of the model, then, could be initially restricted to capturing the conscious world of Lucy's perceptions, representing both a sequence of perceived events, some of them traumatic, and the recollection of those events later on, including their "conversion" into symptoms. The case study mentions many details of Lucy's background and current life, but the specific and significant players in Freud's reconstruction of the case turn out to be few. As a prelude to setting the simulated psychodymanics in motion, I compiled the salient elements from Lucy's reports. The first is the head of the household, a widower, and director of a factory outside of Vienna (the "Director," for short). No less important in Lucy's narrative are the Director's two children, Lucy's pupils. Although each of the two girls would have been separately and complexly represented in Lucy's consciousness, for the purposes of understanding her neurosis it was sufficient to treat them as a single element ("Children"). The several servants working alongside Lucy can be similarly conflated. One other role repeats in important moments in the study, a guest of the director, in one case female and in another male. But as the study unfolds, these two also occupy an interchangeable functional role, abbreviated as "Guest." In Lucy's experience, these key players are not inert, but among their many deeds just one plays a recurrent and significant role in her story: the attempt by each of the guests to kiss the children. These events ("Kissing" in short) will be narrated below. Lucy's "lived world" also features intense feelings. "Love" and "Distress" are prominent, along with the subjective sensations of her chronic rhinitis. Last, but not least, is her perception of burnt pudding and cigar smoke, both the real percepts during moments explicitly identified in the case study, and the later recurrent hallucinations.
In sum, these ten explicit elements within Lucy's unfolding experience compose an initial "alphabet" for Lucynet. Each would have been a complex internal representation, but in pursuit of the minimal model we can regard each as a distinct whole, with varying degrees of prominence from moment to moment in Lucy's history. Different scenes from the saga, then, can be represented by combining subsets of these, as letters might combine to form words. In Lucynet as in Lucy herself, the elements are not isolated, however. Connectionism borrows from a long psychological tradition a single highly general conception of the interaction of ideas, namely, association. Potential "elements" in conscious thought can be activated by associated thoughts, or by inputs from the external world. Which elements are most active at any moment depends on the combination of these influences. In keeping with the radical minimalism of the approach, the activation of any unit at any time was proportional to the sum of its immediately preceding inputs only -- units retained no activation from the previous cycle. By these broad strokes then, we have built an architecture for Lucynet consisting of ten elements -- ten processing units -- suggested by Lucy R., with the potential for interaction both among the elements themselves and from the outside. This architecture is shown in Figure 1. (For a similar approach to modelling memory, see McClelland and Rumelhart 1986.)
[[Please place Figure 1 about here]]
The narrative of learning, and the lessons of trauma
Experience changed Lucy. Whether trivial or traumatic, each episode in the case study left its traces, and so each may be regarded as an occasion of learning. Connectionism captures these experience-driven changes through a process of changing association strength governed by overarching "learning rules." In a network with the simple architecture of Lucynet, we can use a simple form of a very general learning rule known as the delta rule (McClelland and Rumelhart 1986). To a first approximation, the delta rule works as follows: For each cycle of processing, the current level of activity in a unit is compared to the current sum of all inputs to the unit. The connection strengths among the units involved are adjusted so as to reduce this difference. (For details, see Lloyd 1994.) In general, delta rule learning produces networks that are good at associative learning. When Lucynet receives the paired inputs, "Director" and "Children," the delta rule increases the connection strengths between these two units. Subsequently, if Lucynet is presented with either input alone, it will tend to reproduce their combination, the learned association.
The extent of learning with each cycle of operation is governed by a crucial variable, the learning rate coefficient. In connectionist models, this learning rate is kept very low (usually around 0.05) to prevent the learning of one association to interfere with the learning of others. As a consequence, network learning traditionally requires massive repetition of inputs to be learned. With a low learning rate, the same network can learn a number of associations. As a prelude to Lucynet's bumpy ride, I trained the network on a number of background associations that would have characterized Lucy's regular associations in her job. (Some examples: "Director + Children," due to the obvious familial links; "Rhinitis + Distress," reflecting Lucy's chronic complaint; "Children + Love," following Lucy's repeated statements of fondness for her charges.) The network had no trouble learning these associations and reproducing them when partial inputs were provided. In terms of the model, for any given sensory input, Lucynet "recognizes" it (= corresponding unit activation) and experiences "conscious associations" with it (= secondary, weaker activations), both in plausible correspondence to a simple associative psychology conjectured for Lucy herself. During testing, these associations were manifest with the processing cycle immediately following the test input. The network did not need to "settle" through multiple cycles to display its associative pattern completions.
Both Lucy and Lucynet are severely challenged by several subsequent events. Freud identified these episodes as traumatic, although weakly so. To represent these in the changeable web of Lucynet, I tested a crucial hypothesis, the sole modification of connectionist minimalism: An essential concomitant of traumatic experience is learning at an abnormally high learning rate. One main effect of a high learning rate is obvious. Patterns of inputs presented with a high learning rate will be "branded" into a network, swamping prior learning with new associations. In Lucynet, the "traumatic learning" effects occurred with learning rates over 0.3. These effects were increasingly complex, as the subsequent stages of the network will show. Furthermore, at a high learning rate, large effects follow single exposures of patterns to be learned. In this respect, the learning-rate hypothesis corresponds to a fortunate fact about most traumas -- the traumatic event is isolated in its intensity among more ordinary, nontraumatic experiences. The story of Lucy R., then, was reduced in Lucynet to a prelude of unexceptional associative pattern learning, followed by a sequence of single exposure learning trials, where each of Lucy's traumas was modelled by just one exposure, at a "traumatic" high learning rate, to a pattern corresponding to each trauma from the case study.
First, I followed Freud in a slight romantic excess. Lucy's troubles began with falling in love with her boss, the factory director and father of the children in Lucy's care. Freud imagined Lucy's passion beginning with a single meaningful conversation, a sudden and dramatic psychological change. The first pattern for "traumatic learning" was accordingly "Director + Love." The results were as expected. After learning, the background associations involving both units were overwhelmed by the new mutual associations of love. Table 1, row I, displays Freud's account of the moment and its effects, together with their re-creation in Lucynet.
[[Please place table 1 about here.]]
Lucy's love for her employer infuses her reactions to several other events. The first of these is a scene that "crushed her hopes" (118) of a real relationship with the director. The blow came following a visit from a female acquaintance. As the guest prepared to leave, she kissed the two children on the lips. Later, the director shouted at Lucy that such a breech of sanitation was intolerable. If Lucy permitted it to happen again, she would be dismissed. Freud describes this moment as traumatic, the "operative trauma" in the case. The case study is unclear about the immediate effect of the episode (which occurred well before the encounter with Freud). The Lucynet simulation of the event suggests some plausible conjectures. The input for learning is the complex "Director + Guest + Kissing-children + Distress." As expected, subsequent tests after learning showed a pronounced pattern completion effect: Presentation of any one element of the "trauma" led to a pronounced "recall" of the other elements. The sole modulation of this emphatic recall showed a continuing influence of the already established association of units representing the Director and Love. The episode model appears in Row II of Table 1.
DSM IV (American Psychiatric Association, 1994) lists the following phenomenological symptoms for Post-traumatic stress disorder:
B. The traumatic event is persistently reexperienced in one (or more) of the following ways:
(1) recurrent and intrusive distressing recollections of the event, including images, thoughts, or perceptions....
(2) recurrent distressing dreams of the event....
(3) acting or feeling as if the traumatic event were recurring (includes a sense of reliving the experience, illusions, hallucinations, and dissociative flashback episodes)....
(4) intense psychological distress at exposure to internal or external cues that symbolize or resemble an aspect of the traumatic event.
(5) physiological reactivity on exposure to internal or external cues that symbolize or resemble an aspect of the traumatic event.
All of these are plausible elaborations of the pattern completion effects observed in Lucynet. The basic mechanism of PTSD suggested, then, is a mechanism for associative learning and recall operating in a disruptively emphatic way. In Lucynet, the learned complex has no internal structure and every element is equally effective in recreating or "reliving" the experience. As a result, a range of inputs, including trivial reminders of the event, may be sufficient to kindle the whole traumatic pattern.
These observations were very much as expected, given the initial understanding of the delta rule and the effects of increased learning rates. Moreover, the apparent analogy between network learning and PTSD symptoms has been independently noted by Li and Spiegel (1992), who proposed (but did not implement in a model) that trauma be modelled as pattern-completion effects following an emphatic inputs. However, Li and Spiegel did not anticipate the side effects of traumatic learning when traumas compound. For Freud, Lucy, and Lucynet, the more interesting and complex developments lay ahead.
Associations dissociated: Effects of multiple traumas
As it happened, the unpleasantness in Lucy's household recurred. Some weeks after the director's outburst, another guest repeated the attempt to kiss the children. This time the director flared up at the guest, the Chief Accountant at the factory and a regular visitor. To Lucy, however, the scene was "a stab at my heart" (120). One other feature was prominent in Lucy's eventual memory of the scene, the smell of cigar smoke. As in the case study, the Lucynet representation of this scene (which Freud called an "auxiliary trauma") closely parallels the previous traumatic scene, with the noted addition of the smell of cigars. The learned pattern, then, was the complex "Director + Guest + Kissing-children + Cigar-smoke + Distress."
For Freud, it was this scene, rather than the first, that was the origin of Lucy's hysterical symptoms. One might expect that the large overlap between the operative trauma and its auxiliary echo would only reinforce the emphatic learning and pattern completion effects. Yet in Lucynet the responses to the next cycle of traumatic learning departed from the PTSD-like symptoms observed earlier. The first surprise is the nearly complete disappearance of the pattern completion effects (Row III, Table 1). None of the elements of the two traumatic scenes evokes the others. The recall of "distress" and of the main event, the attempt to kiss the children, both seem to vanish. This paradoxical loss of exactly the pattern expected to be retained is the neural network analogue of repression.
Instead of accurate recall of the learned
pattern, the network exhibited an unexpected replacement. To
several input probes, Lucynet responds with the activation of the
unit representing the smell of cigar smoke. Prior to this point
in the simulation, when an input was sent to a single unit of the
network, the strongest response was invariably in that same unit;
we interpreted that response as the perceptual registration of
the input. Now, after compounded trauma, the strongest network
activation no longer corresponds to the input, but is found in
another unit altogether, the "Cigar" representation. An
input is thus converted into a new percept, an activation formed
in the absence of its appropriate input. Lucy experienced a
recurrent hallucination of cigar smoke. Lucynet exhibited an
"inappropriate" activation of its "Cigar"
unit, without the corresponding input. The paradoxical
emergence of a maximal activation without the corresponding input
is the neural network analogue of hallucination.
What is going on here? Freud drew his principle morals from the case study at this point, offering a mechanism to explain the twin observations of repression and symptom formation. Repression begins with a deliberate and conscious effect to banish a painful memory from recall. Memories cannot be erased, however. Instead, they are merely isolated:
This effort to repudiate the hated memory is
thwarted when something in the environment strongly reminds one
of the original trauma. In the case of Lucy R.,
The "convergence" is unbearable,
In short, Freud imagines a dual process:
repression, followed by conversion. His conception of these
processes posits a discrete 'nucleus of thoughts,' explicit
mental representations ("reminiscences") that are
driven from consciousness, but nonetheless reassert themselves in
disguise as symptoms. Because these unbearable thoughts continue
to exist, Freud cheerfully posited an explicit unconscious
'system' to house them. This conception would be elaborated
throughout his career (e.g. Freud 1957 (1915)), but is already
Lucynet models the significant symptoms of Lucy R., and develops those symptoms through a consistent analogue of Lucy's experience in the months before her visit to Freud. But, as figure 1 makes clear, Lucynet utterly lacks the mechanism Freud imagined. Connectionism thus offers a different way of thinking about what occurred in Lucy R., and in cases of compounded trauma in general. Lucynet's "symptoms" are explained by the conjoint effects of the "traumatic learning" of overlapping patterns and the delta learning rule. With a single trauma, the delta rule leads to a pronounced increase in connection strength among the units involved in the traumatic pattern. A single exposure leads to "overlearning," as discussed above. When that same pattern partly repeats, units involved in the pattern receive a flood of input as the external input combines with the massive lateral inputs along the positive, overlearned connections. The delta rule accordingly compensates for this overload by driving down the weights on connections. Since subsequent patterns are also traumatic (that is, are learned at a high learning rate), this inhibitory effect is dramatic. When subsequent patterns partly overlap, the maelstrom of delta rule effects rapidly becomes intractable. In this case, the new element, "Cigar smoke," is exempt from the inhibition affecting the other units, and the network develops the tendency to respond as if that element were present in response to several unrelated inputs.
Lucy's story did not end with the episode just modeled, however, nor was the hallucinated cigar smoke her initial complaint. One more "traumatic" scene followed. In part because of her troubles in her household, Lucy considered quitting her post, but at the price of losing her ties to the children. This conflict of emotions was particularly acute one day as she played with the children just after receiving a letter from her mother, back in Scotland. Just at that moment a pot of pudding on the stove began to burn. Freud reasoned that the conflict of feelings at just that moment was intense enough to constitute a trauma (115), and the smell of burnt pudding its conspicuous marker. As a result, Lucy would hallucinate that smell, her principal complaint henceforth.
Although the pudding only burns once, it seems likely that the emotional tumult that accompanied the smell recurred throughout this period in Lucy's life. If a prior moment of conflict had already been traumatic, then the scene with burnt pudding might have been a repeating, "auxiliary" trauma. In that case, the psychodynamics underlying this symptom would be parallel to the origin of the cigar smoke hallucination, where the first trauma creates massive association, but its repetition massive inhibition and dissociation, save for new elements, which "pop out" as conspicuous new symptoms. Alternatively, Freud proposes that in general an initial trauma and its auxiliary repetition can "coincide," with conversion occurring as an immediate effect (124). In this case, the immediate recall of trauma would itself be traumatic, and even an isolated trauma would become self compounding.
Each of these interpretations suggests different simulations at this stage in Lucynet. Rather than pursue them, however, I kept as close as possible to the case study itself. Freud recounts a single traumatic moment on the theme of leaving the children, a moment also marked, as it happens, by Lucy's rhinitis, and so for Lucynet I input one pattern for traumatic learning: "Children + Rhinitis + Burnt pudding + Distress." Row IV of table 1 shows the results of this next stage of Lucynet traumatic learning. The four elements of the traumatic pattern are bound into a tight associative unit, and the responses parallel the PTSD-like responses also shown in row II. Unlike the cigar smoke hallucination, the smell of burnt pudding does not dominate the response to other inputs. Thus, in the Lucynet framework, it is a pronounced association rather than a clear "hallucination." However, the association is very strong, and elicited by single inputs which Lucy might have encountered routinely: Children, her rhinitis, and her ongoing distress. This simulation also suggests that the memory of the traumatic scene would be readily available, if not intrusive. So it was for Lucy. Freud began his interrogation by asking whether the smell of burnt pudding reminded her of anything, and she was quick to recount just this scene. But that is another story, one of therapy and cure.
Remembrance and catharsis
Having created what may be the world's only neurotic neural network, I felt compelled to restore my creation to full (simulated) mental health. Again, I turned to Freud and Breuer for guidance:
In the Studies, this was called the
"cathartic technique," which Freud would later call
"the immediate precursor of psycho-analysis; and, in spite
of every extension of experience and of every modification of
theory, ... still contained within it as its nucleus" (Freud
1955 (1924): 194).
The network version of catharsis, then, will consist of the re-exposure to the traumatic stimuli. Freud and Breuer stress that the reminiscence of the scene must be accompanied with its original affect. Thus, within the connectionist framework at least some of the original traumatic intensity must accompany the catharsis. I modelled the overall catharsis by re-exposing the network once to each of the traumatic patterns, using a learning rate coefficient set at half that of the original traumatic learning. As in the case study, the patterns were presented in reverse order, as Lucy herself (with Freud's prompting) recounted them.
Catharsis led to cure for Lucynet as for Lucy. Re-exposure to the second auxiliary trauma led to a reduction of associative intensity of inputs to the "burnt pudding" unit (in Lucynet), and a gradual reduction in the hallucination (for Lucy). Re-exposure to the earlier traumas led to a more dramatic reversal. In Lucynet, the "hallucination" of cigar smoke and the accompanying "repression" of the traumatic pattern both disappeared completely, replaced with a normal set of associative links among parts of the traumatic pattern. (For example, the input "Guest" yields activation in "Guest," "Kissing-children," "Cigar smoke," "Director," and "Distress.") For Lucy too, the hallucinated cigar vanished at once. One trauma, however, neither patient nor network could overcome. Both finished their histories with an abiding (if secret) love for their boss. Lucy confessed as much in her last session with Freud; Lucynet showed an implacable two-way association between "Director" and "Love."
To re-experience a trauma one must first remember it. Usually, any number of cues leads to retrieval of a learned pattern. But when a memory is repressed, whether in artificial or human neural networks, many of those associative paths are blocked. How, then, is the repressed pattern recovered? Lucy R. made her way back to the operative trauma by patient association. "At my insistence," Freud wrote, "a picture gradually emerged before her, hesitatingly and piecemeal to being with" (119). So far, Lucynet only displays its immediate response to an input, its "first associations." This made the "repression" of the operative traumatic pattern a barrier to its recall that could not be overcome by any combination of inputs. To further explore this issue, I redesigned Lucynet to simulate a purely internal sequence of thoughts, a simulated "stream of consciousness." To do this, I took the current activation following an initial input, and re-input this activation as a new input. In essence, each pattern of internal activation in the network thus generates its successor, following only the associative paths established between units in the course of the network history. I recorded each of these activation states over ten cycles of recurrent network "reflection."
Examples of the network's changing internal operations are shown in Figure 2. Each panel shows the initial input to the network in the leftmost column, labelled "I." The columns labeled "A" then show ten cycles of response. (The size of the dark squares indicates the magnitude of positive activation. Light squares signify negative activation values, i.e. inhibition.) Cycle 1 is the net's "percept" or immediate response to the input; this is the activation I've interpreted as an analogue to Lucy's state of consciousness at turning points in the case study. Each subsequent cycle is the network's purely internal response to its previous state of activation. For comparison, figure 2 depicts just the evolving responses to input in the "cigar" unit. Until far along in the traumatized training, the net displays a fairly predictable associative "psychology," in which the network settles into a stable state of self-sustaining activation. Two of the background associations (prior to any trauma) are initially apparent: "Director + Guest" (reflecting the director's noted propensity to entertain visitors) and "Director + Guest + Cigar" (following the favorite pastime, noted explicitly by Lucy, of her boss and his male acquaintances). The cigar input "reminds" Lucynet of these associations, and the net settles into a stable recreation of those associative patterns, with echoes of other connections (e.g. "Children + Love").
[[Please place figure 2 about here.]]
Following the overlearning of "Director + Love," Lucynet's soliloquy changes. The cigar still reminds it initially of guests and the director, but once the latter is "in mind," all further thoughts follow the love connection. That stable association is then overthrown by the operative trauma, "Director + Guest + Kissing-children + Distress." As soon as the background associations carry the network toward "Guest" and "Director," it is this most recent "trauma" that dominates, modulated by the prior strong associations between "Director" and "Love."
Up to this point, the network behaves like most connectionist models in that it "settles" into a stable pattern of activation in the absence of further inputs. Simple assumptions about associative psychology and the intense learning of a traumatic pattern explain its behavior. But all this changes in panel D, showing the network's internal processing following the auxiliary trauma, which repeated the preceding trauma with the addition of the smell of cigars. First, the "Cigar" input fails to recreate the traumatic pattern (as discussed above). But rather than settling into a stable response, even without new input the "neurotic" network oscillates between states of excitation and inhibition, each cycle a flashback to an earlier scene in the case study. Significantly, at cycles 4 and 5, the "repressed" traumatic pattern briefly reappears, to be promptly canceled in cycle 6. The compounding of traumatic inputs has not simply erased the history of network learning. That history has radically altered network function, and even left a path for its explicit reconstruction. But the path is a tortuous one. Despite its simplicity, Lucynet captured this aspect of the case study by suggesting the possibility of the recall of repressed content. (Other implications of recurrent processing will be considered below.)
Discussion: From the 90s to the 90s
To summarize, the history of Lucynet follows the case study of Lucy R. by simulating the learning of plausible background associations followed by a series of "traumatic" inputs, learned via single exposures with an abnormally high learning rate coefficient. A second sequence of learning followed the course of therapy, re exposing the network to the "traumatic" input patterns via single exposures at an intermediate learning rate. As a result, the network exhibited robust analogues of three of Freud's most salient observations of Lucy: two olfactory hallucinations and the repression of memory for key episodes of symptom formation. These symptoms emerged and disappeared from the network at moments corresponding to their emergence and disappearance in Freud's chronology of the case. The pattern of symptoms emerged as a side effect of the simulation of Lucy's experience and the operation of the neural network, rather than as a direct result of programming or explicit training to produce these responses.
These are intriguing results given the radical simplicity of Lucynet -- ten units only, governed by a single activation equation and a single learning rule. From ten processing units to the tens of billions in the human brain is a sobering leap. We would not wisely conclude anything about the specific psychology or physiology of Lucy R. or anyone else based on Lucynet. But we can use Lucynet as a heuristic model and foil for theories of hysteria and dissociation. Its simplicity leaves no place for special processes to hide, and thus reveals and questions some widespread assumptions about the mechanisms and etiology of hysteria and its modern descendants.
Perhaps the most useful lesson of Lucynet lies in its deep challenge to what might be called the archival model of memory. By this I mean the conception of memories as fixed records or "reified contents" that can pass in and out of consciousness, and be variously influential or dormant over time. Even in the Studies, Freud clearly conceived of ideas of all sorts in this way:
Once thoughts are reified as special sorts of
fixed objects to be manipulated by the mind, most of the Freudian
mechanics follows as a matter of course. If thoughts exert
influences on behavior and consciousness without themselves
becoming conscious, then they must sometimes exist in the
unconscious (a sort of specialized processing module) and,
moreover, some sort of mental executive must take on the task of
moving thoughts in and out of consciousness, and in an out of
conscious or unconscious play. This way of thinking about
thoughts certainly meshes smoothly with the computational model
of mind that has long dominated cognitive science (Erdelyi 1985).
These days the archival model of memory has been explicitly disavowed by all. Extensive work in cognitive psychology has shown recall to be construction of a memory rather than its retrieval. Connectionists have certainly encouraged this reconception of memory by showing how explicit patterns of activation can be stored implicitly in the form of matrices of connection weights. Memory to a connectionist is a disposition to reform patterns of activation, rather than extract them from some form of storage.
Yet in discussions of psychopathology the archival model and its attendant mechanisms still operate, even if covertly. For example, in his excellent review of dissociative disorders, Kihlstrom (1994) notes that a number of disorders involve failures of recall, but the failures are temporarily or permanently reversible. (This characterizes the DSM IV disorders of dissociative amnesia, dissociative fugue, and dissociative identity disorder.) "Reversible memory disorders are disorders of retrieval; they occur because the individual cannot, at the moment, gain access to memories that have been adequately encoded, and remain available in storage" (Kihlstrom 1994: 379). Here the image of the archive is explicit, although one could alter the terminology to depict storage as a merely dispositional, connectionistic storage. But what Kihlstrom concludes from his observation requires an archival view: "Retrieval, and accessibility, are phenomena of consciousness as they entail bringing available memories into phenomenal awareness" (ibid.). Memories, in short, move in and out of the spotlight of awareness; with the reification of memory comes the reification of a special processor, "consciousness." Kihlstrom and Hoyt (1990) make this explicit as follows:
Thus a conception of memories as fixed records
brings along with it a model of mind in which conscious and
unconscious mental processes can unfold in parallel, passing
thoughts back and forth. The evidence of disintegrated cognition
has suggested to a long line of researchers the existence of
parallel executives. The first to argue along these lines is
probably Plato, who interpreted conflict of the will as evidence
for distinct faculties of mind (Republic, Book. IV). It
persists at the origins of clinical psychology not only in Freud
but also in Janet and James, and in numerous contemporary sources
(e.g. Hilgard). Even the psychopathologists Li and Spiegel, in
their discussion of the import of connectionism for understanding
dissociation, declare a need for the parallel operation of two or
more information processors (Li and Spiegel 1992: 145)
Lucynet gets away with much less. One processor accommodates both Lucynet's preserved "normal" associative processing and its dissociated dislocations. Yet, for all its simplicity, once a certain learning history has transpired, the network ceases to be a passive responder to input stimuli. Figure 2D depicts a new stage in Lucynet evolution and an intriguing moment in connectionist modeling. At this point, patterns vie for expression. While some strut and fret their hour upon the stage, others are (temporarily) heard no more. When various trains of network thought exclude each other, we observe the network analogue of dissociation. But, as always in this study, there is no off stage orchestration. There are just the thoughts on the surface, interacting with each other. The functions of integration or disintegration, making conscious or repressing, are not administered by an agency separated from the thoughts (patterns) themselves.
Instead of a special processing system to monitor and manipulate explicit unconscious representation, through learning Lucynet undergoes widespread changes in the connection weights between conscious elements. These weights define the dispositions of elements to activate one another. They are "unconscious" in the sense that they are part of the implementation of the network rather than its explicitly represented content, analogous to the physiology of synapses in the brain. But this remains a different and less robust conception of the unconscious than that of the archival model.
In addition to its dispositional, connectionist storage of memories, Lucynet exhibits a further break from the archival model: In Lucynet, encoding of new memories alters the encoding of the old. Previously learned patterns change as an immediate side effect of traumatic learning; no special re-enactment of old memories is required. Connectionist modelers usually go to great lengths to prevent the interference of old learning by new, with the goal of accurate reconstruction of discrete learned patterns (Hetherington and Seidenberg 1989; Kortge 1990; Murre 1992). But in Lucynet this interference is exactly the source of both the negative and positive "symptoms" -- dissociation from the overlearned past, and the insertion of a "perceptual hallucination." In most models, interference effects are meaningless noise, but within the guiding framework of a clinical study, Lucynet's wild ride remains interpretable.
The study of Lucy R. has thus become a twice-told tale. Its second telling, as Lucynet, has elided much of the humanity of the first, but it has preserved the main episodes and the main effects noted in Freud's version. The new tale has added a crucial subtext, the fundamental hypothesis that one form of psychogenic pathology can originate when stimulus patterns are subject to intense single exposure learning. But a single episode of "traumatic learning" did not generate the dissociative effects typical of hysteria. For this, our simulated subject needed a history of simulated suffering. Only then did the network become both neurotic and interesting.
Like any story, the new fable of Lucynet is open to many interpretations. Given its simplicity, Lucynet provides no direct evidence about any aspect of human psychology. But it does show something of what is possible on a shoestring. In that spirit, it suggests a few possible morals -- avenues of inquiry worth noting for future clinical research:
1) Psychopathology is a narrative. The scientific emphasis on efficient causation, experimental method, and statistical significance leads to a search for "stories" of pathology in which there are just two episodes, a single cause and its particular effect. Lucynet developed its most revealing syndrome only after a sequence of unique events, each contributing to a complex outcome. In the huge networks that are each of us, every experience and every response reflects the remembrance of things past. Our past experience may not merely provide a general backdrop, but instead contribute in specific ways to otherwise inexplicable responses. Freud, of course, would agree, and most clinicians (but not most insurers). This complicates the understanding of pathology in general, as well as diagnosis and treatment of specific cases. It also threatens the rigor of clinically based science, leading to charges (such as those levelled at psychoanalysis) that it is psuedoscience (e.g. Grünbaum 1984, 1993). Connectionism may offer a middle ground, by allowing for models sensitive to the cumulative effects of personal history, but still constrained by the basic computational capacities of networks. Such models afford further controlled exploration of several variables that may be clinically important.
2) To interpret a clinical narrative, one must understand the perceptual, cognitive, and affective world of the subject. Models like Lucynet are "loose" in several senses. First, they rest on an initial assignment of meanings to network architecture and possible patterns of activation. Second, the traumatic learning is indiscriminate, branding both the central and the trivial elements of a horrific scene into the traumatized memory. But a compounded trauma has the further effect of inhibiting some links within the repeated trauma while enhancing others, leading (in Lucynet, and perhaps in humans) to modifications in subsequent encodings and ultimately to dissociative phenomena. These complex effects in turn depend on the perceptual categories that underlie the recognition of elements as "same" or "different" from one exposure to the next. This category assignment will be sensitive to the ramifying effects of compounded trauma, and to a host of developmental and idiosyncratic differences. If retroactive memory interference occurs in us as well, unravelling a dissociated life narrative may be even more difficult. Event memory traces may not merely be hidden but altered.
3) Connectionism is a multi-level modeling tool. Connectionists often celebrate the "neural inspiration" of their approach, and in recent years have worked toward ever increasing biological realism. The connectionist approach naturally lends itself to the simulation of biological neural networks. But it nonetheless also lends itself to the simulation of other complex phenomena, especially systems subject to multiple simultaneous constraints or internal interactions. Our minds, described at the psychological level as arenas for the interplay of thought, are such systems. It does not matter that Lucynet is biologically unrealistic, or that the delta rule probably does not describe the waxing and waning of synaptic efficacy. What does matter is that the model offers a consistent representation of a domain, so that the model's behavior can be compared with that of entities in the domain. As models become sufficiently general, they become theories of the domain. Lucynet, as a pilot study, cannot pose as a theory of dissociation. But it could indicate a family of psychological models that may ultimately cohere as a theory of some aspect of psychology. That ultimate theory will be no more biological than Lucynet, but none the worse as a theory at its own level.
4) Without a central executive, anything is possible. Just as digital computers suggest (falsely) an image of unfailing and unflappable rationality, connectionist networks project the image of steady pattern-completion and solid, predictable, rote association. They are built to work. But no law enforces this expectation. In fact, these networks are delicate hot-house flowers, reared in the most rigid and contrived learning environments. In the erratic climates of human experience, such nets would fail. Lucynet became a failure as a pattern associator, but an interesting and suggestive one. After his reading of Charcot, Janet, Breuer, and Freud, William James commented:
Between "chaos come again" and the
"neat notions of cognitive and active powers" we find
the vast middle ground of connectionism. In its mechanics and its
poetics we may oneday find a new understanding of both
psychopathology and everyday mental life.
American Psychiatric Association. 1994. Diagnostic and Statistical Manual of Mental Disorders (Fourth Edition). Washington, DC.
Crews, F. 1993. The unknown Freud. New York Review of Books 40(19): 55-65.
Cummins, Robert. 1983.The Nature of Psychological Explanation. MIT Press, Cambridge, MA.
Erdelyi, M. 1985. Psychoanalysis: Freud's Cognitive Psychology San Francisco: Freeman.
Freud, S. 1955 (1924). A short account of psycho-analysis. In Vol. 19 of The Standard Edition of the Complete Psychological Works of Sigmund Freud, ed. James Strachey, pp. 191-209. London: Hogarth.
Freud, S. 1957 (1915). The Unconscious. In Vol. 14 of The Standard Edition of the Complete Psychological Works of Sigmund Freud, ed. James Strachey, pp. 159-217. London: Hogarth.
Freud, S. 1966 (1895). Project for a Scientific Psychology. Vol. 1 of The Standard Edition of the Complete Psychological Works of Sigmund Freud, ed. James Strachey. London: Hogarth.
Freud, S. and Breuer, J. 1955 (1895). Studies on Hysteria. Vol. 2 of The Standard Edition of the Complete Psychological Works of Sigmund Freud, ed. James Strachey. London: Hogarth.
Grünbaum, A. 1984. The foundations of psychoanalysis : a philosophical critique. Berkeley : University of California Press.
Grünbaum, A. 1993. Validation of the clinical theory of psychoanalysis. Madison, CT: International Universities Press.
Hetherington, P. A. and M. Seidenberg. 1989. Is there "catastrophic interference" in connectionist networks? Program of the 11th Annual Conference of the Cognitive Science Society (Erlbaum, Hillsdale, NJ), pp. 26-33.
Kortge, C. A. 1990. Episodic Memory in Connectionist Networks. Program of the 12th Annual Conference of the Cognitive Science Society (Erlbaum, Hillsdale, NJ), pp. 764-771.
James, William. 1982 (1896). Hysteria. In William James on Exceptional Mental States, ed. E. Taylor, pp. 53-72. New York: Scribner's.
Kihstrom, J. 1994. One Hundred Years of Hysteria. In Dissociation: Clinical and Theoretical Perspectives, ed. S. J. Lynn and J. W. Rhue, pp. 365 394. New York: Guilford.
Kihstrom, J. and Hoyt, I. P. 1990. Repression, Dissociation, and Hypnosis. In Repression and dissociation: Implications for personality theory, psychopathology, and health, ed. J. L Singer, pp. 181-208. Chicago: University of Chicago Press.
Li, David and Spiegel, David. 1992. A neural network model of dissociative disorders. Psychiatric Annals 22(3):144-147.
Lloyd, D. 1994. Connectionist Hysteria: Reducing a Freudian Case Study to a Network Model. Philosophy, Psychiatry, and Psychology, 1(2): 69-88.
McClelland, J. and D. Rumelhart. 1986. A Distributed Model of Human Learning and Memory. In Parallel Distributed Processing: Explorations in the Microstructure of Cognition: Volume 2: Psychological and Biological Models, ed. J. McClelland, D. Rumelhart, and the PDP Research Group, pp. 170-216. Cambridge, MA: MIT Press,
Murre, J. 1992. Effects of pattern presentation on interference in back propagation networks. Program of the 14th Annual Conference of the Cognitive Science Society (Erlbaum, Hillsdale, NJ), pp. 54-59.
Shank, R. C. And Abelson, R. P. 1977. Scripts, Plans, Goals, and Understanding. Hillsdale, NJ: Erlbaum.
Wegman, C. 1985. Psychoanalysis and Cognitive Psychology . London: Academic.