No CrossRef data available.
Article contents
Reverse engineering the structure of cognitive mechanisms
Published online by Cambridge University Press: 25 August 2011
Abstract
Describing a cognitive system at a mechanistic level requires an engineering task analysis. This involves identifying the task and developing models of possible solutions. Evolutionary psychology and Bayesian modeling make complimentary contributions: Evolutionary psychology suggests the types of tasks that human brains were designed to solve, while Bayesian modeling provides a rigorous description of possible computational solutions to such problems.
Information
- Type
- Open Peer Commentary
- Information
- Copyright
- Copyright © Cambridge University Press 2011
References
Cosmides, L. & Tooby, J. (1987) From evolution to behavior: Evolutionary psychology as the missing link. In: The latest on the best: Essays on evolution and optimality, ed. Dupre, J., pp. 277–306. MIT Press.Google Scholar
Marr, D. (1982) Vision: A computational investigation into the human representation and processing of visual information. W. H. Freeman.Google Scholar
Perrigo, G., Belvin, L. & Vom Saal, F. S. (1991) Individual variation in the neural timing of infanticide and parental behavior in male house mice. Physiology and Behavior
50:287–96.CrossRefGoogle ScholarPubMed
Perrigo, G., Belvin, L. & Vom Saal, F. S. (1992) Time and sex in the male mouse: Temporal regulation of infanticide and parental behavior. Chronobiology International
9:421–33.CrossRefGoogle ScholarPubMed
Tooby, J., Cosmides, L., Sell, A., Lieberman, D. & Sznycer, D. (2008) Internal regulatory variables and the design of human motivation: A computational and evolutionary approach. In: Handbook of approach and avoidance motivation, ed. Elliot, A., pp. 251–71. Erlbaum.Google Scholar
Because of their mathematical formalism, Bayesian models of cognition have the potential to infuse greater rigor into psychological models of how the mind works. Any theoretical framework committed to specifying (1) the class of cues that a mechanism is a sensitive to, (2) the operations it performs in response to those cues, and (3) the resultant outputs, is to be heartily welcomed into the theoretical toolbox of psychology.
Jones & Love (J&L) argue that, to be successful, Bayesian modelers should increase their focus on a mechanistic level of analysis, and use the examples of behaviorism and evolutionary psychology to warn them against the pitfalls of theoretical approaches that ignore psychological mechanisms and instead move directly from behavior to the environment. In the case of evolutionary psychology, this critique is simply mistaken. In fact, the field was founded specifically in response to previous evolutionary approaches, such as ethology, that ignored this middle level of analysis (e.g., Cosmides & Tooby Reference Cosmides, Tooby and Dupre1987). The goal of evolutionary psychology is the same as any branch of cognitive science: to describe the information-processing structure of psychological mechanisms. What is distinct about evolutionary psychology is that principles of natural selection are used to predict the structure of cognitive mechanisms. These models generate testable predictions that can be adjudicated by empirical data.
The history of psychology suggests that well-specified task analyses (Marr Reference Marr1982) are the most tractable way of reverse engineering the structure of cognitive mechanisms. As J&L discuss, the challenge for any psychologist is to (1) identify the task being solved, and (2) develop models of possible solutions. Through this lens, evolutionary psychology and Bayesian modeling make complimentary contributions. Evolutionary psychology, properly applied, is a deductive framework for generating predictions about the types of tasks cognitive mechanisms were designed to solve. This constrains the possibility space for the structure of a cognitive mechanism – what class of cues mechanisms are likely to use and what their resultant output and criterion for success should be. It rests on the premise that natural selection builds deterministic cognitive mechanisms that take as inputs aspects of the world that were invariant over phylogenetic time and generate outputs that would have led to the intergenerational differential reproduction of such systems. It is therefore a way to deductively generate hypotheses about the existence of previously unknown cognitive mechanisms. What evolutionary psychology is not – even in principle – is a description of any particular engineering solution. In contrast, Bayesian modeling is a description of an engineering solution: Cognitive mechanisms whose function requires holding and updating probabilities will – constraints aside – behave according to Bayes’ theorem. What Bayesian modeling is not – even in principle – is a way to generate hypotheses or predictions about the range of cues cognitive systems use and their criteria for success.
Natural selection builds cognitive mechanisms around phylogenetic invariances
Organisms' cognitive mechanisms reflect the dynamics of multiple generations of individuals interacting with recurrent features of the natural environment (i.e., phylogenetic dynamics not visible within an individual lifetime). For example, after copulation with a female, a male house mouse will commit infanticide on any pups born for the duration of the typical mouse gestational period, after which point they will rear any pups born; manipulations demonstrated that males achieve this by tracking the number of light/dark cycles (Perrigo et al. Reference Perrigo, Belvin and Vom Saal1991; Reference Perrigo, Belvin and Vom Saal1992). The cognitive mechanisms in the male mouse that mediate this relationship between light/dark cycles and killing versus caring behaviors are a result of the dynamics of differential reproductive success over multiple generations. Invariant relationships in the world – the duration of light/dark cycles in a natural terrestrial environment, that copulation leads to offspring, the duration of gestation, and so forth – are “seen” by natural selection, which in turn engineers biological mechanisms that instantiate input/output relationships. In this example, not only are the input cues based around intergenerational invariances, but the generated outputs are those which would lead to differential reproductive success within the context of those intergenerational invariances (i.e., mechanisms that discriminately kill or rear pups as a function of actuarial relatedness will do differentially better over multiple generations than mechanisms that do not).
As this example demonstrates, differential reproductive success (i.e., natural selection) operating over phylogenetic invariances determines input/output relationships in cognitive systems (see Tooby et al. [Reference Tooby, Cosmides, Sell, Lieberman, Sznycer and Elliot2008] for examples of using the deductive logic of phylogenetic invariances to predict and test novel cognitive mechanisms in humans). Of course, once a task is identified and a relevant mechanism proposed, the computational structure of that mechanism must still be described. Any one particular computational engineering solution is not entailed by the fact that natural selection designed a certain cognitive mechanism – in principle, there are many possible engineering solutions. In some cases, the computational solution to handling a particular set of invariances will be a Bayesian system. Integrating a phylogenetic perspective (in addition to an ontogenetic one) can provide Bayesian modelers with clear, deductive ways to determine the hypothesis space for a computational system and to set priors.
Going forward: Engineering task analyses
Historical accident aside, Bayesian modeling and evolutionary psychology are not in fact alternative approaches to understanding psychology. Rather, both make necessary but distinct contributions to the process of reverse engineering the mind at a mechanistic level. We are confident that both evolutionary psychology and Bayesian modeling could productively pool their efforts. Evolutionary psychology can provide the framing of task analyses – descriptions of the problem and tasks that cognitive systems must in principle solve. Bayesian models of cognition can provide rigorous, mathematical descriptions of certain types of engineering solutions. We look forward to a time when psychologists choose ecologically valid task analyses and posit fully mechanistic accounts of how the solution to those problems could be implemented by a fully mechanistic system without trying to shoe-horn each reverse engineering task analysis into any common overarching meta-theoretical framework. In the future, we hope there are no evolutionary psychologists or Bayesian modelers, just psychologists who reverse engineer the mind at a mechanistic level, using any and all deductive theoretical tools at their disposal.