Rahnev & Denison (R&D) make the convincing case, with a lengthy review of examples of suboptimal perceptual performance, that broad claims of optimality in perceptual decision making are false. We applaud the overall spirit of R&D's analysis. They make the point that the pursuit of demonstrations of optimality is unproductive because “optimality is … only well-defined in the context of a set of specific assumptions, rendering general statements about the optimality (or suboptimality) of human perceptual decisions meaningless” (sect. 4.1, para. 4). We concur also with the authors’ call to test models, something that many of the cited researchers are already doing, rather than telling optimality-based “just-so stories.” R&D's proposed approach is likely to do part of the work in avoiding some of the pitfalls laid out by Bowers and Davis (Reference Bowers and Davis2012a).
In our view, R&D's main points should be taken even further, in at least two respects. First, R&D's arguments apply well beyond the realm of perceptual decision making, with relevance for other areas of mind and brain science. The article mentions that cue combination studies have been fundamental to the view that perception is optimal. In addition to the given examples of cue combination across multiple senses, or sensorimotor integration, similar claims of optimality have been made, across development, for a wide range of tasks involving judgments of stimuli from memory (Cicchini et al. Reference Cicchini, Arrighi, Cecchetti, Giusti and Burr2012; Duffy et al. Reference Duffy, Huttenlocher and Crawford2006; Huttenlocher et al. Reference Huttenlocher, Hedges and Vevea2000; Sciutti et al. Reference Sciutti, Burr, Saracco, Sandini and Gori2015; see also Petzschner et al. Reference Petzschner, Glasauer and Stephan2015). Claims of optimality are widespread in this literature, and just as in perceptual decision making, because “optimality is often the hypothesized outcome … findings of suboptimality may be underreported or underemphasized” (R&D, sect. 3.7.2, para. 1). For example, this literature emphasizes models developed to explain central tendency biases in the reproduction of temporal and spatial magnitudes and characterizes those biases as optimal, yet there are multiple instances of data that violate the assumptions of these models (e.g., Barth et al. Reference Barth, Lesser, Taggart and Slusser2015; see also Crawford & Duffy Reference Crawford and Duffy2010; Duffy & Smith Reference Duffy and Smith2017).
Second, seeking to explain suboptimal, as well as optimal, behavior may not be sufficient to avoid the just-so story dilemma. R&D focus on the idea that broad claims about perceptual optimality are meaningless and untestable, but this objection does not apply only to optimality per se: It applies to broad claims about perception as Bayesian inference more generally. As many critics have pointed out, both within and beyond the context of perceptual decision making, “in almost all published cases, some Bayesian model can be found that approximates or parallels human performance” (Anderson Reference Anderson2015, p. 280). To use an example mentioned briefly by R&D, some visual illusions are particularly difficult to explain in the context of theories of optimal perception. Usually, illusory surfaces or contours are perceived when the visual system needs to “explain” the disappearance or occlusion of an image element. The percept of an illusory occluding surface (a surface that is not defined by contrast edges) is often thought to occur because it is the most probable explanation of that disappearance. But in one class of dynamic displays, an illusory occluding surface is perceived even though a clearly visible occluding surface is already present. The disappearance of image elements is already fully explained by the presence of the visible occluder, yet our perceptual systems construct a second highly improbable occluding surface – the illusory one – generating a redundant “explanation.” The phenomenon has therefore been taken as evidence against claims of visual contour synthesis as a case of rational or Bayesian inference (Anderson et al. Reference Anderson, O'Vari and Barth2011). However, as other researchers have pointed out, despite the fact that “the presence of the spurious contours results from higher-level constraints than those captured by a single Bayesian model concerned only with contour completion,” the phenomenon could still be explained within a broader Bayesian view of the brain (Fleming Reference Fleming2011, p. R261). This is one of many examples of the ultimately untestable nature of claims about perception as Bayesian inference (for a deeper analysis of theoretical challenges to the Bayesian program in visual perception, see Anderson [Reference Anderson2015]).
Are we trying to make a general argument against the adoption of Bayesian perspectives? No. We argue for conceptual clarity regarding the status of overarching theoretical perspectives, and against the common practice of having it both ways. Circularity is unavoidable if researchers work entirely within Bayesian frameworks while, at the same time, specific findings are taken as evidence that Bayesian perception or cognition are in play. That is, researchers who choose to take a broadly Bayesian perspective as a matter of a priori theoretical preference, such that only Bayesian explanations of observed phenomena are entertained, must resist the temptation to claim that correspondences between their data and their models provide evidence for their broad perspective. Some researchers are careful to limit their claims, presenting supporting evidence only for particular models. Others are not, creating highly flexible Bayesian models to fit patterns of behavior and finding – unsurprisingly – that their data match the predictions of these models and then using that match to make strong general claims about Bayesian brains. But the perspective itself – although it may be a productive environment for thinking about how the mind might work – is not falsifiable, and so evidentiary claims at the level of that perspective are unsupportable.
There is something to be lost when any one perspective dominates the literature. The problem is not just limiting inquiry to the chase for the aesthetic ideal of optimality; it is also concerning if we limit inquiry to the use of models that fall within particular theoretical frameworks. Testing very different kinds of models, rather than working within a single theoretical perspective, is important too. The authors conclude that researchers “should aim for their models to capture all of the systematic weirdness of human behavior rather than preserve an aesthetic ideal” (R&D, sect. 6, para. 1). The quest to capture that systematic weirdness will undoubtedly benefit from theoretical heterodoxy in the field.
Rahnev & Denison (R&D) make the convincing case, with a lengthy review of examples of suboptimal perceptual performance, that broad claims of optimality in perceptual decision making are false. We applaud the overall spirit of R&D's analysis. They make the point that the pursuit of demonstrations of optimality is unproductive because “optimality is … only well-defined in the context of a set of specific assumptions, rendering general statements about the optimality (or suboptimality) of human perceptual decisions meaningless” (sect. 4.1, para. 4). We concur also with the authors’ call to test models, something that many of the cited researchers are already doing, rather than telling optimality-based “just-so stories.” R&D's proposed approach is likely to do part of the work in avoiding some of the pitfalls laid out by Bowers and Davis (Reference Bowers and Davis2012a).
In our view, R&D's main points should be taken even further, in at least two respects. First, R&D's arguments apply well beyond the realm of perceptual decision making, with relevance for other areas of mind and brain science. The article mentions that cue combination studies have been fundamental to the view that perception is optimal. In addition to the given examples of cue combination across multiple senses, or sensorimotor integration, similar claims of optimality have been made, across development, for a wide range of tasks involving judgments of stimuli from memory (Cicchini et al. Reference Cicchini, Arrighi, Cecchetti, Giusti and Burr2012; Duffy et al. Reference Duffy, Huttenlocher and Crawford2006; Huttenlocher et al. Reference Huttenlocher, Hedges and Vevea2000; Sciutti et al. Reference Sciutti, Burr, Saracco, Sandini and Gori2015; see also Petzschner et al. Reference Petzschner, Glasauer and Stephan2015). Claims of optimality are widespread in this literature, and just as in perceptual decision making, because “optimality is often the hypothesized outcome … findings of suboptimality may be underreported or underemphasized” (R&D, sect. 3.7.2, para. 1). For example, this literature emphasizes models developed to explain central tendency biases in the reproduction of temporal and spatial magnitudes and characterizes those biases as optimal, yet there are multiple instances of data that violate the assumptions of these models (e.g., Barth et al. Reference Barth, Lesser, Taggart and Slusser2015; see also Crawford & Duffy Reference Crawford and Duffy2010; Duffy & Smith Reference Duffy and Smith2017).
Second, seeking to explain suboptimal, as well as optimal, behavior may not be sufficient to avoid the just-so story dilemma. R&D focus on the idea that broad claims about perceptual optimality are meaningless and untestable, but this objection does not apply only to optimality per se: It applies to broad claims about perception as Bayesian inference more generally. As many critics have pointed out, both within and beyond the context of perceptual decision making, “in almost all published cases, some Bayesian model can be found that approximates or parallels human performance” (Anderson Reference Anderson2015, p. 280). To use an example mentioned briefly by R&D, some visual illusions are particularly difficult to explain in the context of theories of optimal perception. Usually, illusory surfaces or contours are perceived when the visual system needs to “explain” the disappearance or occlusion of an image element. The percept of an illusory occluding surface (a surface that is not defined by contrast edges) is often thought to occur because it is the most probable explanation of that disappearance. But in one class of dynamic displays, an illusory occluding surface is perceived even though a clearly visible occluding surface is already present. The disappearance of image elements is already fully explained by the presence of the visible occluder, yet our perceptual systems construct a second highly improbable occluding surface – the illusory one – generating a redundant “explanation.” The phenomenon has therefore been taken as evidence against claims of visual contour synthesis as a case of rational or Bayesian inference (Anderson et al. Reference Anderson, O'Vari and Barth2011). However, as other researchers have pointed out, despite the fact that “the presence of the spurious contours results from higher-level constraints than those captured by a single Bayesian model concerned only with contour completion,” the phenomenon could still be explained within a broader Bayesian view of the brain (Fleming Reference Fleming2011, p. R261). This is one of many examples of the ultimately untestable nature of claims about perception as Bayesian inference (for a deeper analysis of theoretical challenges to the Bayesian program in visual perception, see Anderson [Reference Anderson2015]).
Are we trying to make a general argument against the adoption of Bayesian perspectives? No. We argue for conceptual clarity regarding the status of overarching theoretical perspectives, and against the common practice of having it both ways. Circularity is unavoidable if researchers work entirely within Bayesian frameworks while, at the same time, specific findings are taken as evidence that Bayesian perception or cognition are in play. That is, researchers who choose to take a broadly Bayesian perspective as a matter of a priori theoretical preference, such that only Bayesian explanations of observed phenomena are entertained, must resist the temptation to claim that correspondences between their data and their models provide evidence for their broad perspective. Some researchers are careful to limit their claims, presenting supporting evidence only for particular models. Others are not, creating highly flexible Bayesian models to fit patterns of behavior and finding – unsurprisingly – that their data match the predictions of these models and then using that match to make strong general claims about Bayesian brains. But the perspective itself – although it may be a productive environment for thinking about how the mind might work – is not falsifiable, and so evidentiary claims at the level of that perspective are unsupportable.
There is something to be lost when any one perspective dominates the literature. The problem is not just limiting inquiry to the chase for the aesthetic ideal of optimality; it is also concerning if we limit inquiry to the use of models that fall within particular theoretical frameworks. Testing very different kinds of models, rather than working within a single theoretical perspective, is important too. The authors conclude that researchers “should aim for their models to capture all of the systematic weirdness of human behavior rather than preserve an aesthetic ideal” (R&D, sect. 6, para. 1). The quest to capture that systematic weirdness will undoubtedly benefit from theoretical heterodoxy in the field.