4.2 Motivations for Computational Modeling
4.2.1 Typical Uses of Computational Models
In cognitive neuroscience, a typical computational model proposes, explores, and/or tests a hypothesized neural explanation for a psychological effect or neurological disorder. In order to do this, a computational model is built that is structured vaguely in accordance with some known neuroanatomical or physiological facts or hypotheses. In what sense a model needs to share the structure of the target system is open to several interpretations. A successful computational model should reproduce in its results the psychological effect or neurological disorder aimed at, and do a better job of it than any competing models. This is referred to in the field as simulating the data. (Simulation takes on quite another meaning in the philosophical literature on computational modeling.)
Many models compare two (or more) conditions: one where the model is ‘damaged’ and one where it is not. The damage might take the form of connections between parts of the system being disabled, a module being removed, or a change being made to the function that determines the activity of some parts; any of these may be rough methods of simulating neural damage. The goal is to reproduce the behavior of both patients and controls. If the model succeeds at all of this, a tentative conclusion is drawn suggesting that the neural system might work in the same way as the computational model.
If the computational model produces interesting results or suggests a neural structure that is not yet known to exist in the target system, this is considered a prediction of the
model. Confirmation of these predictions, whether behavioral or neural, would make the model more of a success, especially if the new finding is surprising or counterintuitive.
An often mentioned benefit of computational models is that compared to ‘verbal’ models (models of how a system is thought to work in the form of diagrams, charts, or verbal descrip- tions), computational models must be completely specified, so conjectures about whether the model works as intended are settled by running the program.
This picture of a typical computational model’s methodology is informed by a wide survey of models selected from cognitive neuroscience textbooks, the top journals in the field, and citation lists. The particular papers that I describe in detail below to illustrate this picture were selected by searching for ‘computational model cognitive neuroscience’ on GoogleScholar, and selecting all the papers and chapters about attention models published in 2000 or later in the first two pages of hits. Although I focus on models of attention, computational models of other cognitive phenomena follow the same pattern (seeO’Loughlin and Thagard (2000), Coltheart et al. (2001) for well-cited models of autism and reading errors, for example).
Amos(2000) presents a model of performance on the Wisconsin Card Sort Task (WCST), a measure of executive attention, that compares patients with schizophrenia, Parkinson’s disease, Huntington’s disease, and healthy controls. Amos criticizes two previous models for not taking into account anatomical details, and one for not implementing a simulation of patient data. He says,
A network constrained by neuroanatomy and patient data might engender a finer level of description of processes involved in the performance of the WCST, and of the information processing performed in the frontal cortex and other areas of the cortico-basal loops (Amos 2000).
Anatomical details of frontal cortex and the cortico-basal loops are incorporated in the model, including columnar organization, excitatory projections to the striatum, and topographically organized projections from the substantia nigra and globus pallidus. Physiological research is also incorporated, for example by modeling dopamine dysfunction in schizophrenia as a reduction in the gain on affected neurons’ activation functions. Although these and other biologically realistic details are built in, the model is still highly simplified in terms of the connections between brain areas implemented, the number of ‘units’ in each modeled brain
area, and the details included in these units.
Damage to the system is simulated in several ways for the various conditions considered, and to explore alternative hypothesized models for these conditions, for example by increas- ing bias against firing in frontal units, decreasing gain in the striatal module, reducing the output of striatal neurons, and adjusting noise parameters. The patterns of errors produced by the model under the various damage conditions are compared to behavioral results from the patient populations. The hypotheses corresponding to patterns of errors that better match the behavioral data are taken to be confirmed. Amos (2000) says that, “Simulation of quantitative data from patient populations may help differentiate the explanatory value of these approaches,” which I take to mean that implementing competing hypotheses in computational models, and trying to reproduce the patient data in some detail is a way of deciding which of the hypotheses provides the best explanation of the data.
Amos suggests several predictions arising from his model and gives some details as to how these might be tested in single-unit studies on monkeys, and behavioral studies on schizophrenic, Parkinson’s and Huntington’s patients. In addition, Amos suggests a possi- ble link between the three diseases considered; they may overlap in their symptomatology because of a common mechanism. Several times Amos mentions that he is looking for “in- formation processing” mechanisms, although it is never made clear what exactly he means by this.
O’Reilly and Frank (2006) also describe a computational model of executive attention that is “based on the prefrontal cortex and basal ganglia.” Their goal is a biologically plausi- ble model that performs working memory tasks as well as previous, less biologically plausible models. They suggest that their model has “direct implications for understanding executive dysfunction in neurological disorders such as attention deficit-hyperactivity disorder (ADHD) and Parkinson’s disease” (O’Reilly and Frank 2006), so again there is interest in accounting for both normal psychological effects and neurological disorders. Furthermore, the model generates testable predictions, and the authors claim their model can test hypotheses about the causes of neurological symptoms: “we think the model can explicitly test the implica- tions of striatal dopamine dysfunction in producing cognitive deficits in conditions such as Parkinson’s disease and ADHD” (O’Reilly and Frank 2006).
This model is rather more focused on the computational problems it needs to solve than that of Amos (2000), but still makes a point of getting both the biological details (the anatomical connections between frontal cortex and basal ganglia structures) and the behavioral data (performance of various groups on working memory tasks) right. They go into relatively more detail about certain aspects of the system, like the timing of the signals passing between the different parts of the circuit, but again the biologically plausibility is limited; the model “omits many biological details of the real system” (O’Reilly and Frank 2006, 31). Again there are comparisons to other models of the same behavioral data, and the selling points of this model are described both in terms of it matching the behavioral data more closely than other models, and in terms of it being more biologically plausible.
A telling indication of their general approach to integration is the following claim, “Be- cause the PBWM model represents a level of modeling intermediate between detailed biolog- ical models and powerful, abstract cognitive and computational models, it has the potential to build important bridges between these disparate levels of analysis” (O’Reilly and Frank 2006). The virtue of a model being ‘intermediate’ in this way is a topic we will return to later.
De Pisapia et al. (2008) give an overview of computational models of attention in their textbook chapter. They highlight many of the same methodological points. Biological plau- sibility is valued, but only up to a point. They note that most computational models of visual attention share the same organization “which follows at least coarsely the struc- ture and organization of the visual perceptual system” (De Pisapia et al. 2008), including modules representing brain areas V1, PP and IT, hypercolumns, refractory periods, local inhibitory connections, center-surround receptive fields, and so on. They also highlight integration of multiple explanatory levels “from single-cell neurophysiology to observable behavior” (De Pisapia et al. 2008).
The method I have described as being typical for computational models in cognitive neuroscience involves a combination of inferences. The logic of tendencies seems to be involved in the concern with roughly sharing some structural properties with brain systems, as it might be inferred that systems with like structure should behave in similar ways. Inferences to like causes from like effects seem also to be involved in concluding that a
simulation whose results match behavioral data the most closely are most likely to capture the underlying mechanisms giving rise to the behavior. What seems to be going on is inference to the best explanation; models are compared to one another to see which most closely matches both the underlying structure and the behavioral data, and the one that would make for the best explanation of the evidence is preferred, at least until a better competitor model comes along. What Amos (2000) may be getting at when he talks about constraints coming from both neuroanatomy and patient data is that there is a sort of squeezing from both sides, or an inferential pincer movement being used. The methodology depends on both sets of constraints being used together to narrow down the space of possible models.
In terms of the kind of explanation being sought in this sample of studies, there is much talk of underlying mechanisms. Uncovering the neural mechanisms that give rise to psychological phenomena or neurological conditions is what these studies are after. But the aim is very obviously not to get all the nitty-gritty details just right. The mechanisms they are looking for may be quite general. A nod is made to the value of unification in explanation in Amos’s appeal to a common mechanism shared by schizophrenia, Parkinson’s and Huntington’s diseases. O’Reilly et al. talk about their model being intermediate between detailed biology and abstract computation. In later sections I will discuss in more detail why models with this intermediate status between biological plausibility and abstract mechanism might be preferred by scientists.
This sample is not intended to argue that this methodology is the only one employed in computational modeling in cognitive neuroscience. In fact plenty of computational models in cognitive neuroscience have other aims entirely, and correspondingly different methods. This plurality of methods and aims will become important shortly. The particular sort of model I describe here is significant, because computational explanations of cognitive phenomena do tend to take this form, and as will become clear in Section 4.3, questions have been raised as to how this sort of intermediate model can help explain cognition.