Tuesday, April 29, 2014

Foundations of Computational Psychiatry


Computational psychiatry is a new field that seeks to identify cognitive (dys)functions that are indicative of mental illness. Here is a definition from Quentin Huys in the Encyclopedia of Computational Neuroscience:
Computational Psychiatry is a heterogeneous field at the intersection of computational neuroscience and psychiatry. Incorporating methods from psychiatry, psychology, neuroscience, behavioural economics and machine learning, computational psychiatry focuses on building mathematical models of neural or cognitive phenomena relevant to psychiatric diseases. The models span a wide range - from biologically detailed models of neurons or networks to abstract models describing high-level cognitive abilities of an organism. Psychiatric diseases are conceptualized as either an extreme of normal function, or as a consequence of alterations in parts of the model.

As in computational neuroscience more generally, the building of models forces key concepts to be made concrete and hidden assumptions to be made explicit. One critical functions of these models in the setting of psychiatry are their ability to bridge between low-level biological and high-level cognitive features. While many neurobiological alterations are known, the exclusively atheoretical focus of standard psychiatric nosology on high-level symptoms has as yet prevented an integration of these bodies of knowledge.
This is an interesting new field. Here are two articles, one recent and one from a few years ago, that can serve as a foundation for entry into this new, eclectic field.

Full Citation:
Wallace R. (2014) Foundations of computational psychiatry: a mathematical treatment of embodied cognition and its disorders. PeerJ PrePrints 2:e356v1

Foundations of computational psychiatry: a mathematical treatment of embodied cognition and its disorders

Rodrick Wallace


The Data Rate Theorem that establishes a formal linkage between real-time linear control theory and information theory carries deep implications for the study of embodied cognition and its dysfunctions across human, machine, and composite man-machine cockpit entities. The stabilization of such cognition is a dynamic process deeply intertwined with it, constituting, in a sense, the riverbanks directing the flow of a stream of generalized consciousness. The fundamental role of culture in human life has long been understood. Here we argue that, not only is culture, in the words of the evolutionary anthropologist Robert Boyd, 'as much a part of human biology as the enamel on our teeth', but that it is an important part of the disorders of embodied cognition that different communities socially construct as mental illness. A shift in perspective on such illness is badly needed, from naive geneticism and a simplistic 'biomarker' focus to a broad cognitive theory that recognizes the central roles of culture, socioeconomic structure, their histories and their dynamics, in human development. Unfortunately, stabilizing analogs to cultural influence are unavailable to high-order real-time machine and cockpit systems that, increasingly, will be tasked with control of critical enterprises ranging from communication, power, and travel networks to nuclear and chemical reactors, organized deadly force, and so on.

1 Introduction

Varela, Thompson and Rosch (1991), in their pioneering study The Embodied Mind: Cognitive Science and Human Experience, asserted that the world is portrayed and determined by mutual interaction between the physiology of an organism, its sensorimotor circuitry, and the environment. The essential point, in their view, being the inherent structural coupling of brain-body-world. Lively debate has followed and continues (e.g., Clark, 1998; M. Wilson, 2002; A. Wilson and S. Golonka, 2013). As Wilson and Golonka put it,
The most exciting hypothesis in cognitive science right now is the theory that cognition is embodied... [T]he implications of embodiment are... radical... If cognition can span the brain, body, and environment, then the `states of mind' of disembodied cognitive science won't exist to be modified. Cognition will instead be an extended system assembled from a broad variety of resources... [that] can span brain, body, and environment...
[We] treat perception-action problems and language problems as the same kind of thing... Linguistic information is a task resource in exactly the same way as perceptual information... Our behavior emerges from a pool of potential task resources that include the body, the environment and... the brain...
In the same year as the work by Varela et al., Brooks (1991) explored analogous ideas in robotics;
Artificial intelligence research has foundered on the issue of representation. When intelligence is approached in an incremental manner, with strict reliance on interfacing to the real world through perception and action, reliance on representation disappears... The fundamental decomposition of the intelligent system is not into independent information processing units which must interface with each other via representation. Instead, the intelligent system is decomposed into independent and parallel activity producers which all interface directly to the world through perception and action... everything is both central and peripheral...
Barrett and Henzi (2005) summarize matters as follows:
...[C]ognition is `situated' and `distributed'. Cognition is not limited by the `skin and skull' of the individual... but uses resources and materials in the environment... The dynamic social interactions of primates... can be investigated as cognitive processes in themselves... A distributed approach... considers all cognitive processes to emerge from the interactions between individuals, and between individuals and the world.
We will explore below how this picture, of the recruitment of disparate resources into shifting, temporary coalitions in real time to address challenges and opportunities, represents a significant extension of the Baars model of consciousness.
It is possible to study cognition and embodiment through the newly-uncovered Data Rate Theorem that, for the first time, formally relates control theory and information theory, and to include as well the needed regulation and stabilization mechanisms in a unitary construct that must interpenetrate in a similar manner.

Natural cognitive systems operate at all scales and levels of organization of biological process (e.g., Wallace, 2012, 2014a). The failure of low level biological cognition in humans is often expressed through early onset of the intractable chronic diseases of senescence (e.g., Wallace and Wallace, 2010, 2013). Failure of high order cognition in humans has been the subject of intensive scientific study for over two hundred years, with little if any consensus. As Johnson-Laird et al. (2006) put it,
Current knowledge about psychological illnesses is comparable to the medical understanding of epidemics in the early 19th Century. Physicians realized that cholera, for example, was a specific disease, which killed about a third of the people whom it infected. What they disagreed about was the cause, the pathology, and the communication of the disease. Similarly, most medical professionals these days realize that psychological illnesses occur... but they disagree about their cause and pathology.
As the media chatter surrounding the release of the latest official US nosology of mental disorders - the so-called `DSM-V' - indicates, this may be something an understatement. Indeed, the entire enterprise of the Diagnostic and Statistical Manual of Mental Disorders has been characterized as `prescientific' (e.g., Gilbert, 2001). Atmanspacher (2006), for example, argues that formal theory of high-level cognition is itself at a point like that of physics 400 years ago, with the basic entities and the relations between them yet to be determined. Further complications arise via the overwhelming in uence of culture on both mental process and its dysfunction (e.g., Heine, 2001; Kleinman and Cohen 1997), something to which we will return. Chapter 5 of Wallace and Wallace (2013) provides a more detailed exploration.

The stabilization and regulation of high order cognition may be as complex as such cognition itself.

Some simplification, however, is possible. Cognition can be described in terms of a sophisticated real-time feedback between interior and exterior, necessarily constrained, as Dretske (1994) has noted, by certain asymptotic limit theorems of probability:
Communication theory can be interpreted as telling one something important about the conditions that are needed for the transmission of information as ordinarily understood, about what it takes for the transmission of semantic information. This has tempted people... to exploit [information theory] in semantic and cognitive studies...
...Unless there is a statistically reliable channel of communication between [a source and a receiver]... no signal can carry semantic information... [thus] the channel over which the [semantic] signal arrives [must satisfy] the appropriate statistical constraints of information theory.
Recent intersection of that theory with the formalisms of real-time feedback systems - control theory - may provide insight into matters of embodied cognition and the parallel synergistic problem of embodied regulation and control. Here, we extend that work and apply the resulting conceptual model toward formally characterizing the unitary structural coupling of brain-body-world, a coupling in which, for humans, culture is a central mode. In the process, we will explore dynamic statistical models that can be fitted to data.
We will find, after some considerable development, results recognizably similar to those of Verduzco-Flores et al. (2012). They, using a detailed neural network model, observed that
[Certain network] changes result in a set of dynamics which may be associated with cognitive symptoms associated with different neuropathologies, particularly epilepsy, schizophrenia, and obsessive compulsive disorders... [S]ymptoms in these disorders may arise from similar or the same general mechanisms...[and] these pathological dynamics may form a set of overlapping states within the normal network function..[related] to observed associations between different pathologies.
Apparently, these are robust observations across a variety of cognitive systems.
Read the whole article.

Here is an older article on computational psychiatry, which has a different feel to me. I suspect this is such a new field of study that it is still figuring out what it is.

Full Citation:
Montague, PR, Dolan, RJ, Friston, KJ, and Dayan, P. (2011, Dec 14). Computational Psychiatry.
Trends Cogn Sci.; 16(1): 72–80. doi: 10.1016/j.tics.2011.11.018

Computational psychiatry

P. Read Montague, Raymond J. Dolan, Karl J. Friston, and Peter Dayan


Computational ideas pervade many areas of science and have an integrative explanatory role in neuroscience and cognitive science. However, computational depictions of cognitive function have had surprisingly little impact on the way we assess mental illness because diseases of the mind have not been systematically conceptualized in computational terms. Here, we outline goals and nascent efforts in the new field of computational psychiatry, which seeks to characterize mental dysfunction in terms of aberrant computations over multiple scales. We highlight early efforts in this area that employ reinforcement learning and game theoretic frameworks to elucidate decision-making in health and disease. Looking forwards, we emphasize a need for theory development and large-scale computational phenotyping in human subjects.

The explanatory gap

The idea of biological psychiatry seems simple and compelling: the brain is the organ that generates, sustains and supports mental function, and modern psychiatry seeks the biological basis of mental illnesses. This approach has been a primary driver behind the development of generations of anti-psychotic, anti-depressant, and anti-anxiety drugs that enjoy widespread clinical use. Despite this progress, biological psychiatry and neuroscience face an enormous explanatory gap. This gap represents a lack of appropriate intermediate levels of description that bind ideas articulated at the molecular level to those expressed at the level of descriptive clinical entities, such as schizophrenia, depression and anxiety. In general, we lack a sufficient understanding of human cognition (and cognitive phenotypes) to provide a bridge between the molecular and the phenomenological. This is reflected in questions and concerns regarding the classification of psychiatric diseases themselves, notably, each time the Diagnostic and Statistical Manual of Mental Disorders (DSM) of the American Psychiatric Association is revised [1].

While multiple causes are likely to account for the current state of affairs, one contributor to this gap is the (almost) unreasonable effectiveness of psychotropic medication. These medications are of great benefit to a substantial number of patients; however, our understanding of why they work on mental function remains rudimentary. For example, receptors are understood as molecular motifs (encoded by genes) that shuttle information from one cellular site to another. Receptor ligands, whose blockade or activation relieves psychiatric symptoms, furnished a kind of conceptual leap that seemed to obviate the need to account for the numerous layers of representation intervening between receptor function and behavioral change. This, in turn, spawned explanations of mental phenomena in simplistic terms that invoked a direct mapping from receptor activation to complex changes in mental status. We are all participants in this state of affairs, since symptom relief in severe mental disease is sufficient from a clinical perspective, irrespective of whether there are models that connect underlying biological phenomena to the damaged mental function. A medication that relieves or removes symptoms in a large population of subjects is unquestionably of great utility, even if the explanation for why it works is lacking. However, significant gaps in the effectiveness of medications for different mental illness mean we should look to advances in modern neuroscience and cognitive science to deliver more.

We believe that advances in human neuroscience can bridge parts of the explanatory gap. One area where there has been substantial progress is in the field of decision-making. Aberrant decision-making is central to the majority of psychiatric conditions and this provides a unique opportunity for progress. It is the computational revolution in cognitive neuroscience that underpins this opportunity and argues strongly for the application of computational approaches to psychiatry. This is the basis of computational psychiatry [24] (Figure 1). In this article, we consider this emerging field and outline central challenges for the immediate future.

An external file that holds a picture, illustration, etc.
Object name is nihms345286f1.jpg
Figure 1 - Components of Computational Psychiatry.

Contrasting mathematical and computational modeling

Mathematical modeling

To define computational modeling, we must first distinguish it from its close cousin, mathematical or biophysical modeling. Mathematical modeling provides a quantitative expression for natural phenomena. This may involve building multi-level (unifying) reductive accounts of natural phenomena. The reductions involve explanatory models at one level of description that are based on models at finer levels, and are ubiquitous in everything from treatments of action potentials [5] (see also [6] for a broader view) to the dynamical activity of populations of recurrently connected neurons [7]. Biophysical realism, however, is a harsh taskmaster, particularly in the face of incomplete or sparse data. For example, in humans, there seems to be little point in building a biophysically detailed model of the dendrite of single neurons if one can only measure synaptic responses averaged over millions of neurons and billions of synapses using functional magnetic resonance imaging (fMRI) or electroencephalography (EEG).

Biophysical modeling is important for elucidating key relationships in a hugely complex system [8] and thus predicting the possible effects of therapeutic interventions (see [9] for an example using dynamic causal modeling). For example, it is well known that critical mechanisms within neuromodulatory systems, such as dopamine, serotonin, norepinephrine and acetylcholine, are subject to intricate patterns of feedback and interactive control, with autoreceptors regulating the activity of the very neurons that release neuromodulators. Moreover, this feedback often includes the effects of one neuromodulator (e.g., serotonin) on the release and impact of others (e.g., dopamine) [10]. These neuromodulators are implicated in many psychiatric and neurological conditions. The fact that they play key roles in so many critical functions may explain the fact, if not the nature, of this exquisite regulation. It is the complexity of these interactions that invites biophysical modeling and simulation, for instance, to predict the effect of medication with known effects on receptors or uptake mechanisms. Moreover, the capacity to perform fast biophysical simulations is essential for evidence-based model comparison using empirical data [11] and the exploration of emergent behaviors (e.g., [12]). Simulation has become vital to vast areas of science and it will be central in computational psychiatry. Mathematical predictions based on real neural and biophysical data are important; however, they are not equivalent to a computational account of mental or neural function.

Computational modeling

Computational modeling seeks normative computational accounts of neural and cognitive function. Such accounts start from the premise that the brain solves computational problems and indeed has evolved to do so. One of the pioneers of computing theory, Alan Turing, conceived of mental function in exactly this fashion – the mind was cast as specific patterns of information processing supported by a particular kind of hardware (the brain) [13]. This notion implies key constraints on mental phenomena – in particular constraints on computational complexity that limit the power of any device, neural or mechanical, to solve a wide range of problems [14]. This idea is commonplace today but in the 1930s the idea of computation and its limits underwent a revolution [1517] (see also [18]).

Currently, computational accounts of elements of mental and neural function exist, and in each case, typically some constraint is found that guides the discovery of the computational model. Some of the most important constraints come from optimality assumptions – the idea that the brain is organized to maximize or minimize quantities of external and internal importance (e.g., [6,19]). One set of optimality constraints emerges naturally from behaviors that support survival, such as foraging for food or responding appropriately to prospects of danger [20]. A wide range of ideas, proofs, methods and algorithms for executing such behaviors can be found in many fields, including engineering, economics, operations research, control theory, statistics, artificial intelligence and computer science. In fact, these fields provide a formal foundation for the interpretation of many cognitive and neural phenomena [6]. This foundation can span important levels of description, for instance, offering accounts of the representational semantics of the population activity of neurons [21] or of the firing of neuromodulatory neurons in the context of tasks involving predictions of reward [2226]. This type of computational modeling can thus provide one explanatory framework for the reductive mathematical modeling discussed above.

Computational modeling in decision-making

The field of decision-making has been a particular target for computational modeling. Decision-making involves the accumulation of evidence associated with the utilities of possible options and then the choice of one of them, given the evidence. Decision-making problems in natural environments are extremely complex. One difficulty arises from the balance models must strike between built-in information acquired over the course of evolution about the nature of the decision-making environment (ultimate constraints) versus what can be learned over the course of moment-to-moment experience (proximate constraints). A second difficulty arises because of the inherent computational complexity of the problem: certain types of optimal decision-making appear intractable for any computational system. This fact motivates the search for approximations that underlie mechanisms actually used in animals. Reinforcement learning is one area where such approximations have been used to guide the discovery of neural and behavioral mechanisms. Box 1 provides a brief description of the modern view of neural reinforcement learning.
Box 1. Reinforcement learning

Reinforcement learning (RL) is a field, partly spawned by mathematical psychology, that spans artificial intelligence, operations research, statistics and control theory (for a good introductory account of RL, see [89]). RL addresses how systems of any sort, be they artificial or natural, can learn to gain rewards and avoid punishments in what might be very complicated environments, involving states (such as locations in a maze) and transitions between states. The field of neural RL maps RL concepts and algorithms onto aspects of the neural substrate of affective decision-making [90,91]. One important feature of this framework is that the majority of its models can be derived from a normative model of how an agent ‘should’ behave under some explicit notion of what that agent is trying to optimize [89].

Conventional and neural RL include two very broad classes of method: model-based and model-free. Model-based RL involves building a statistical model of the environment (a form of cognitive map; see [92]) and then using it to (i) choose actions based on predicted outcomes and (ii) improve predictions by optimizing the model. Acquiring such models from experience can be enhanced by sophisticated prior expectations (a facet that we relate to the phenomena of learned helplessness). In other words, an agent significantly enhances the models it can build based on experience if it already starts with a good characterization of its environment. In turn, these models enable moment-to-moment prediction and planning. Except in very simple environments, prediction and planning consume enormous memory and computational resources – a fact has inspired much work on approximations and the search for biological work-arounds.

Model-free RL involves learning exactly the same predictions and preferences as model-based RL, but without building a model. Instead, model-free RL learns predictions about the environment by enforcing a strong consistency constraint: successive predictions about the same future outcomes should be the same. Actions are chosen based on the simple principle that actions which lead to better predicted outcomes are preferred. Model-free RL imposes much lower demands on computation and memory because it depends on past learning rather than present inference. However, this makes it less flexible to changes in the environment.

The conceptual differences between model-based and model-free RL suggest that correlates can be sought in real-world neural and behavioral data. There are ample results from animal and human experiments to suggest that both model-free and model-based RL systems exist in partially distinct regions of the brain [67,9397] and that there is a rich panoply of competitive and cooperative interactions between them [67]. Model-free RL has a particularly close association with the activity of the dopamine neuromodulatory system, especially in the context of appetitive outcomes and predictions.

Finally, model-based and model-free RL are both instrumental in the sense that actions are chosen because of their consequences [94]. Animals are also endowed with extremely sophisticated Pavlovian controllers (see main text), where outcomes and predictions of those outcomes directly elicit a set of species-typical choices apparently not under voluntary control. One important example related to predictions of future negative outcomes is behavioral inhibition (learning not to do something), which may be related to serotonin [51].
Many psychiatric conditions are associated not only with abnormal subjective states, such as moods, but also with aberrant decisions. Patients make choices: in depression, not to explore; in obsessive compulsive disorder, to repeat endlessly a behavior (such as hand-washing) that has no apparent basis in rational fact (such as having dirty hands); in addiction, to seek and take a drug, despite explicitly acknowledging the damage that follows. Key to the initial form of computational psychiatry is the premise that, if the psychology and neurobiology of normative decision-making can be characterized and parameterized via a multi-level computational framework, it will be possible to understand the many ways in which decision-making can go wrong. However, we should first consider an important earlier tradition of modeling in psychiatry.

Early connectionist models of mental dysfunction

There is an old idea in brain science, namely, that complex functions emerge from networked interactions of relatively simple parts [27,28]. In the brain, the most conspicuous physical substrates for this idea are the networks of neurons connected by synapses. This perspective has been termed ‘connectionism’. One modern expression of connectionism began with the work of Rumelhart, McClelland and the parallel distributed processing research group [29] (but now see [30]), which applied this approach to both brain and cognition in the early and mid-1980s, building upon the earlier pioneering work [27,28]. The basic concept underlying connectionism involves taking simple, neuron-like, units and connecting in them in ways that are either biologically plausible based on brain data or capable of performing important cognitive or behavioral functions. At approximately the same time and parallel to this work, three key publications emerged from physicists John Hopfield and David Tank, which showed how a connectionist-like network can have properties equivalent to those pertaining to the dynamics of a physical system [3133]. Inspired by Hopfield’s work and the seminal (and still classic) work of Stuart and Donald Geman on Gibbs sampling and Bayesian approaches to image analysis, Hinton and Sejnowski [34] showed that probabilistic activation in simple units could perform a sophisticated Bayesian style of inference. Collectively, this work addressed memory states, constraint satisfaction, pattern recognition and a host of other cognitive functions [29], thus suggesting that these models might aid in understanding mental disease.

Through the 1990s, connectionist models turned their sights on psychopathologies, such as schizophrenia [3539]. These models primarily addressed issues related to cognitive control and neuromodulation [3538], with a particular focus on neural systems that could support these functions [4044]. These and other models offered plausible solutions for how networks of neurons could implement functions, such as cognitive control and memory, and offered new abstractions for how such functions go awry in specific pathologies. This work leans heavily on the neurally-plausible aspect of connectionist models, a feature that now finds more biological support, as neuroscience has produced enormous amounts of new data that can be fit into such frameworks [4244].

Recent efforts toward computational characterization of mental dysfunction

In this section, we review recent efforts to develop and test computational models of mental dysfunction and to extract behavioral phenotypes relevant for building computationally-principled models of mental disease. The examples discussed are intended to provide insights into healthy mental function but in a fashion designed to inform the diagnosis and treatment of mental disease. Along with the pioneering earlier studies [3540], there have been recent treatments and reports of work along these lines on schizophrenia [3,45,46], addiction [47], Parkinson’s disease, Tourette’s syndrome, and attention-deficit hyperactivity disorder [3]. Here, we concentrate on two areas that have not been recently reviewed in this context, namely depression and autism.

The efforts discussed here are now collectively blossoming into programmatic efforts in computational psychiatry (for example, the joint initiative of the Max Planck Society and University College London: Computational Psychiatry and Aging Research). It is our opinion that such efforts must reach further and strive to extract normative computational accounts of healthy and pathological cognition useful for building predictive models of individuals. Consequently, we emphasize for computational psychiatry the goal of extracting computational principles around which human cognition and its supporting biological apparatus is organized. Achieving this goal will require new types of phenotyping approaches, in which computational parameters are estimated (neurally and behaviorally) from human subjects and used to inform the models. This type of large-scale computational phenotyping of human behavior does not yet exist.

Reinforcement learning (RL) models of mood disorders and anxiety

Box 1 notes three different, albeit interacting, control systems within the context of RL: model-based, model-free, and Pavlovian. Model-based and model-free systems link the choice of actions directly to affective consequences. The Pavlovian system determines involuntary actions on the basis of predictions of outcomes, whether or not deployed actions are actually appropriate for gaining or avoiding those outcomes. Pavlovian control appears completely automated in this description. However, it is known that other brain systems can interact with Pavlovian control, hence, it is at this level that such control can be sensitive to ongoing valuations in other parts of the brain.

These types of controllers and their interactions have been the subject of computational modeling in the context of mood disorders, especially depression [4,4850]. First, let us consider the role of serotonin in clinical depression. In many patients, one effective treatment involves the use of a selective serotonin reuptake inhibitor (SSRI), which prolongs the action of serotonin at target sites. Data from animals suggests that serotonin release is involved in (learned) behavioral inhibition [5053], associated with the prediction of aversive outcomes [54,55]. Computational modeling inspired by these data suggests that serotonin’s role in behavioral inhibition may reflect a Pavlovian effect: subjects do not have to learn explicitly what (not) to do in the face of possible future trouble. This effect could be called the ‘serotonergic crutch’. Problems with the operation of this crutch can lead to behavior in which poor choices are made because they have not been learned to be inappropriate. In this framework, punishments are experienced or imagined even if the choices concern internal trains of thought rather than external events [50]. Restoration of the crutch is considered to improve matters again. The logic here is that the more an individual’s behavior is determined in a Pavlovian manner, the more devastating is the likely consequence of any problem with the serotonergic crutch. This is an account of vulnerability (analogous to the incentive sensitization theory of drug addiction; see [5658]).

Conversely, model-based RL has been used to capture another feature of some forms of anxiety and depression: learned helplessness [5961]. Animals can be made helpless when provided with uncontrollable rewards as well as uncontrollable punishments [62] and, thus, learning that their actions do not consistently predict outcomes. In these experiments, one way to demonstrate the onset of learned helplessness is to show that the animals do not explore or try to escape when placed in new environments (e.g. [63]).

A natural computational account is to treat the helplessness training in the first part of learned helplessness as inducing a prior probability distribution over possible future environments, indicating that the animal can expect to have little influence over its fate, that is, little controllability. This hypothesis is based on the expectation that related environments have similar properties. Exploration in a new environment is only worthwhile only if it is expected that good outcomes can be reliably achieved given appropriate actions. Thus, a prior belief implying that the environment is unlikely to afford substantial controllability will discourage exploration. Prior distributions are under active examination in Bayesian approaches to cognitive science and are offering substantial explanations for a broad range of developmental and adult behaviors [64]. Only model-based RL is capable of incorporating such rich priors, even though model-free control can be induced to behave in similar ways by simpler mechanisms [65,66]. This computational interpretation of uncontrollability provides a new way to understand the role that environments can play in etiology. It also provides a way of formalizing the complex interaction between model-based, model-free and Pavlovian systems, when not only might one controller directly influence the training signal of the other controllers [67] but also the very experience other controllers require in order to learn their own predictions or courses of action.

Using games to phenotype autism spectrum disorder

A defining feature of human cognition is the capacity to model and understand the intentions (and emotions) of other humans. This extends to an ability to forecast into the near-term future, for example, how someone else will feel should they experience a consequence of an action that we might take. Sophisticated capacities such as these lie at the heart of our ability to cooperate, compete and communicate with others. One of the defining features of autism spectrum disorder (ASD) is a diminished capacity for socio-emotional reciprocity – the social back-and-forth engagement associated with all human interaction [1,6870]. Recent modeling and neuroimaging work has used two-agent interactions, typically in the form of some game, to parameterize and probe this social give-and-take [7181]. This work, along with other efforts [8285], has collectively launched a computational neuroscience perspective on inter-personal exchange – a first step toward identifying computational phenotypes in human interactions that are underwritten by both behavioral and neural responses.

Game theory is the study of mathematical models of interacting rational agents. It is used in many domains and in recent years has been increasingly applied to common behavioral interactions in humans. Two game-theoretic approaches have recently been used to probe ASD and other psychopathological populations directly: the stag hunt game and the multi-round trust game (Figure 2). Although the behavioral probes are different, the two games share the feature that it is advantageous for a human player to make inferences about their partner’s likely mental state during the game. These inferences are recursive: my model of you incorporates my model of your model of me, and so on. Both approaches have built computational models around this central idea of recursion [78,79], thereby furnishing component computations required for healthy human exchange.

An external file that holds a picture, illustration, etc.
Object name is nihms345286f2.jpg
Figure 2 - Economic games requiring theory-of-mind modeling
Yoshida et al. [7981] used the stag hunt game (Figure 2a) to probe mental state inferences in ASD versus control subjects. The stag hunt game is a classic two-player game (in this case involving a human player and a computer agent), where players can cooperate to hunt and acquire high yield stags, or act alone and hunt low yield rabbits. The model developed by Yoshida et al. [79] used the human player’s observed behavior to estimate the sophistication level (depth of recursion) of their inference about the computer agent’s beliefs (theory of mind). This estimate is necessary if the human player is to cooperate successfully with the computer agent – the human must believe that the agent believes that the human will also cooperate, and so on.

Behavioral results that exploited this model pointed to a higher probability for a theory-of-mind model (versus fixed-strategy) for control subjects. The opposite was true for ASD subjects (~78% probability for fixed-strategy). However, as one might expect, there was heterogeneity in these estimates, with some ASD subjects (n=5) displaying higher probability for the theory-of-mind model compared to a fixed-strategy model (n=12). Intriguingly, ASD subjects with a higher probability for a fixed-strategy model showed higher ratings on two ASD rating scales (ADI-R, ASDI). These results are preliminary and the sample of ASD subjects small. However, the crucial point is that the model allows for a principled parameterization of important cognitive components (e.g., depth of recursion in modeling one’s partner). The use of such a model provides a way to formalize the cognitive components of ASD in computational terms. By collecting much more normative data, this type of approach could serve to differentiate ASD along these newly defined computational dimensions to improve diagnosis, guide other modes of investigation and help tailor treatments.

The multi-round trust game has also been used to probe a range of psychopathologic populations including ASD and borderline personality disorder [2,75,77]. The game is a sequential fairness game involving reciprocation, where performance is determined by whether players think through the impact of their actions on their partner. In the game, a proposer (called the investor) is endowed with $20 and chooses to send some fraction I to their partner. This fraction is tripled (to 3*I) on the way to the responder (called the trustee), who then chooses to send back some fraction of the tripled amount. Subjects play 10 rounds and know this beforehand. Cooperation earns both players the most money. Even when playing with an anonymous partner, investors do send money, a fact that challenges rational agent accounts of such exchanges. One way to conceptualize this willingness to send money was proposed by Fehr and Schmidt [86], who suggested that in such a social setting a player’s utility for money depends on the fairness of the split across the two players. Based on this model of fair exchange between humans, Ray and colleagues developed a Bayesian model of how one player ‘mentalizes’ the impact of their actions (money split with partner) on their partner [78]. The key feature is for each player to observe monetary exchanges with their partner and estimate in a Bayesian manner the ‘fairness type’ of their partner, that is, the degree to which the partner is sensitive to an inequitable split.

This model was able to ‘type’ players reliably from 8 rounds of monetary exchange in the game. These types can be used to seek type-specific (fairness sensitivity) neural correlates. More importantly, the model can be used to phenotype individuals according to computational parameters important in this simple game-theoretic model of human exchange. This is an important new possibility. Using this same game, Koshelev and colleagues showed that healthy investors playing with a range of psychopathological groups in the trustee role can be clustered in a manner that reflects that type of psychopathology acting as the trustee [87]. This model used a Bayesian clustering approach to observations of the healthy investors’ behavior as induced by interactions with different psychopathology groups. These preliminary results suggest that parameters extracted from staged (normative) game-theoretic exchanges could be used profitably as a new phenotyping tool for humans, where the phenotypes are defined by computational parameters extracted using models.

Computational phenotyping of human cognitive function

Computational models of human mental function present more general possibilities for producing new and useful human phenotypes. These phenotypes can then structure the search for genetic and neural contributions to healthy and diseased cognition. We do not expect such an approach to supplant current descriptive nosologies; instead, they will be an adjunct, where the nature of the computational characterization offers a new lexicon for understanding mental function in humans. Moreover, this approach can start with humans, define a computational phenotype, seek neural and genetic correlates of this phenotype and then turn to animal models for deeper biological study.

Under the restricted decision-making landscape that we have painted, RL models provide a natural example of a type of computational model that could be used in such phenotyping. Moreover, we sketched briefly how game-theoretic probes also allow for new forms of computational modeling and hence new ways to computationally phenotype humans. Through their built-in principles of operation and notions of optimal performance, RL models provide constraints that help bridge the aforementioned gap between molecular and behavioral levels of description. However, the behavioral underpinning of these models is extremely shallow at present, especially in human subjects. As suggested by the examples above, the estimation and use of computational variables, such as these, will require new kinds of behavioral probes, combined with an ever-evolving capacity to make neural measurements in healthy human brains. Not only is better phenotyping through the development of new probes needed, but also unprecedented levels of phenotyping of cognitive function. Many of the best ideas about mental performance and function derive primarily from studies in other species. While these animal models have been strikingly successful at uncovering the biology underlying learning, memory and behavioral choice, the human behavioral ‘software’ is likely to be significantly different in important ways that the probes will need to capture. Large-scale computational phenotyping will require radical levels of openness across scientific disciplines and successful models for data exchange and data sharing.

Concluding remarks

If the computational approaches we have outlined turn out to be effective in psychiatry, then what might one expect? The large-scale behavioral phenotyping project sketched above involves substantial aspects of data analysis and computational modeling. The aim of the data analysis will be to link precise elements of the models to measurable aspects of behavior and to molecular and neural substrates that can be independently measured. A strong likelihood here is that the models will offer a set of categories for dysfunction that are related to, but different from, existing notions of disease and this will lead to a need for translation.

Although we did not focus on them here, there are also implications for mathematical modeling. A simulation-based account of measurable brain dynamics, anatomical pathways and brain regions could be expected, equipped with visualization and analysis methods to help make sense of the output. The ultimate hope is for a detailed, multi-level, model that allows prediction of the effects of malfunctions and manipulations. However, making this sufficiently accurate at the scales that matter for cognition and behavior is a long way off. One critical, though as yet unproven, possibility is that a computational understanding will provide its own kind of short-circuit, with, for instance, rules of self-organization of neural elements based on achieving particular computational endpoints, thereby removing the requirement for detailed specification.

Finally, the most pressing requirement is for training. Broad and deep skills across cognitive neuroscience, computational neuroscience, cellular and molecular neuroscience, pharmacology, neurology, and psychiatry itself, in addition to computer science and engineering, are required for the emergence of the richly interdisciplinary field of computational psychiatry. Optimistically, how to achieve this may become clearer as thoughts mature about restructuring education to achieve breadth across the brain-related clinical disciplines of neurology and psychiatry [88].


Cognitive phenotype - a phenotype is a measureable trait of an organism. Although easy to state in this manner, the idea of a phenotype can become subtle and contentious. Phenotypes include different morphology, biochemical cascades, neural connection patterns, behavioral patterns and so on. 
Phenotypic variation is a term used to refer to those variations in some trait on which natural selection could act. A cognitive phenotype is a pattern of cognitive functioning in some domain that could be used to classify styles of cognition. By analogy, variations in cognitive phenotypes would be subject to natural selection.
Computational phenotyping - a computational phenotype is a measurable behavioral or neural type defined in terms of some computational model. By analogy with other phenotypes, a computational phenotype should show variation across individuals and natural selection could act on this variation. Large-scale computational phenotyping in humans has not been carried out; therefore, the ultimate utility of this idea has not been rigorously tested.
Game theory - the study of mathematical models of interactions between rational agents.
Instrumental controller - instrumental conditioning is the process by which reward and punishment are used in a contingent fashion to increase or decrease the likelihood that some behavior will occur again in the future. An instrumental controller is one whose control over behavior can be conditioned in exactly the same fashion. It is an operational term used in the reinforcement learning approach to motivated behavior to refer to any controller whose influence over behavior shows the dependence on rewards and punishments typical of instrumental conditioning.
Neuromodulatory systems - systems of neurons that project to broad regions of target neural tissue to modulate subsequent neural responses in those regions. Neuromodulatory systems typically have cell bodies situated in the brainstem and basal forebrain and deliver neurotransmitters, such as serotonin, dopamine, acetylcholine and norepinephrine, to target regions. They are called modulatory because their impact is typically much longer-lasting than fast synaptic effects mediated by glutamate and they are much more widely distributed.
Pavlovian controller - an operational name for a behavioral controller that is Pavlovian in the normal psychological use of this term – that is, the controller mediates involuntary responses to situations or stimuli. Pavlovian control can be demonstrated behaviorally and modern work is focused on identifying the neural substrates that contribute to this function.
Serotonin - a neuromodulator common to many neurons in the raphe nuclei. Serotonin has a presumed role in clinical depression because of the efficacy of medications that selectively block its reuptake into neurons after its release from synaptic terminals (so-called SSRI’s – selective serotonin reuptake inhibitors).

No comments: