There are two main interrelated perspectives in the measurement of perception – “outer psychophysics” with its behavioral responses and “inner psychophysics” with its neurobiological responses to stimulation (Fechner Reference Fechner1882). The measurements in outer psychophysics depend on the objective stimuli and the neural processes dealing with the signals from the stimuli. These neural processes as the basis for perceptual judgments not only depend on the input signals but, as Rahnev & Denison (R&D) rightfully seem to accept, they are also heavily influenced by a host of factors unrelated to sensory signal-to-noise ratio. There are difficult problems with valid setting of the posterior probability and knowing sufficiently the priors related actual factors. Determining whether observers’ decisions are optimal necessitates sufficiently detailed specification of all likelihood function, prior, cost function, and decision rule (LPCD) components. It must be stressed that substantial individual differences in the behavioral phenotypes and endophenotypes impacting perceptual judgments make a major source of obstacles for developing a powerful general model founded on the optimal decision rule.
Inner psychophysics variability as dependent on common genetic variability cannot be accounted for by any strict generalized formal models expected to be valid for the majority of experimental subjects. Given the huge number of genes implicated in neurobiological mechanism–based perceptual responses, any general models of perceptual decisions that would support the optimality theory are impossible. Indeed, even when only a few genes with variants associated with different serotonergic or dopaminergic endophenotypes are used as experimental independent variables, metacontrast masking in two-alternative discrimination shows considerable individual differences (Maksimov et al. Reference Maksimov, Vaht, Harro and Bachmann2013; Reference Maksimov, Vaht, Harro and Bachmann2015a; Reference Maksimov, Vaht, Murd, Harro and Bachmann2015b). Considering that there are a multitude of genes with effects expressed also in cholinergic, noradrenergic, glutamatergic, and other functions (obviously impacting perceptual behavior) and that already any optional relatively small observers’ sample can invoke covert neurobiological variability with even more varying perceptual effects, any model aspiring to be general cannot be meaningfully established. Therefore, the recommendation by R&D to test more specific models with specific priors and individual limitations (as the sources that corrupt decisions) seems right. As the number of free parameters in genetic variability is very large, the task of distilling and grouping individualized endophenotypal priors as independent experimental variables is not quick and easy. But it may have importance considering that many genes implicated in perceptual behavior are also the so-called risk genes in terms of their relation to health-related vulnerability.
R&D argue that dissociations and low degree of correspondence between confidence and accuracy make general statements about the optimality or suboptimality of perceptual decisions meaningless; similar problems are related to perceptual biases, illusions, and improbabilities (sect. 3.5.2 and 3.8.1). However, the concepts of noise and bias are ambiguous when we move from general information-theoretic perceptual decision models to real psychophysiological processes and mechanisms and want to use these models in consciousness research. For example, associative learning–based expectations tend to produce hallucination-like illusions such that observers experience auditory or visual stimuli that are objectively absent (Aru & Bachmann Reference Aru and Bachmann2017; Mack et al. Reference Mack, Erol, Clarke and Bert2016; Powers et al. Reference Powers, Mathys and Corlett2017). Importantly, the proneness to such illusions is individually variable. Bayesian prior related top-down effects can produce perceptual illusions or hallucinations that are technically noise (and must be included in noise distribution), but they are at the same time phenomenal perceptual experiences taken as veridical by the observers. Hence, this “technical noise” becomes an important factor of real awareness based behavior. On the other hand, noise can be interpreted traditionally as the factor countering sensitivity simply by decreasing the signal-to-noise ratio without any illusory non-veridical experiences. Similarly, hallucinatory experiences are technically indications of bias (type 1 bias) but are at the same time factors determining behavioral choices. The other version of the bias effect relates to its impact on decision without illusory experiences of the imperative stimuli being involved (type 2 bias). In the case of illusory non-veridical percepts, high confidence ratings are expected, but in the case of no illusory experience (but low signal-to-noise ratio), low confidence ratings are expected. Both the concept of noise and the concept of bias in the traditional perceptual decision models used for advocating optimality are underspecified or even misleading in the light of experimental data on consciousness contents and qualia of perception, including illusions and hallucinations. Behaviorally incorrect perceptual responses can be subjectively right, leading to high confidence.
With regard to the domain of decision rules, R&D note that observers may not consider all information relevant to the decision. Moreover, in most of the stimulation that is present for evaluation by the observer, there are several cues on which the decision can be based. When perceptual decision models (and also methods of scaling subjective experiences such as the Perceptual Awareness Scale [PAS]; Sandberg et al. Reference Sandberg, Timmermans, Overgaard and Cleeremans2010) use too general, underspecified evaluation criteria such as level of confidence or subjective clarity, difficulties arise in interpreting the results and explaining the underlying mechanisms and processes. The same behavioral outcome (e.g., successful discrimination or recognition) can be founded on different aspects of subjective contents of perception as different criterion contents (Jannati & Di Lollo Reference Jannati and Di Lollo2012; Kahneman Reference Kahneman1968). For behaviorist models of perceptual decisions, this may not be a problem, but for models to be used for consciousness studies, the excess generality of scales and criteria is a shortcoming. The cure would be development and use of specific rating scales for different subjective aspects of the same stimulation. This would be useful also for accounting for the effects of cue dissociation, which is an alternative source of ambiguities in addition to redundancy effects from cue combination.
To conclude, I tend to agree with R&D in that general statements about the optimality or suboptimality of perceptual decisions are meaningless unless detailed enough observer models will be put together.
There are two main interrelated perspectives in the measurement of perception – “outer psychophysics” with its behavioral responses and “inner psychophysics” with its neurobiological responses to stimulation (Fechner Reference Fechner1882). The measurements in outer psychophysics depend on the objective stimuli and the neural processes dealing with the signals from the stimuli. These neural processes as the basis for perceptual judgments not only depend on the input signals but, as Rahnev & Denison (R&D) rightfully seem to accept, they are also heavily influenced by a host of factors unrelated to sensory signal-to-noise ratio. There are difficult problems with valid setting of the posterior probability and knowing sufficiently the priors related actual factors. Determining whether observers’ decisions are optimal necessitates sufficiently detailed specification of all likelihood function, prior, cost function, and decision rule (LPCD) components. It must be stressed that substantial individual differences in the behavioral phenotypes and endophenotypes impacting perceptual judgments make a major source of obstacles for developing a powerful general model founded on the optimal decision rule.
Inner psychophysics variability as dependent on common genetic variability cannot be accounted for by any strict generalized formal models expected to be valid for the majority of experimental subjects. Given the huge number of genes implicated in neurobiological mechanism–based perceptual responses, any general models of perceptual decisions that would support the optimality theory are impossible. Indeed, even when only a few genes with variants associated with different serotonergic or dopaminergic endophenotypes are used as experimental independent variables, metacontrast masking in two-alternative discrimination shows considerable individual differences (Maksimov et al. Reference Maksimov, Vaht, Harro and Bachmann2013; Reference Maksimov, Vaht, Harro and Bachmann2015a; Reference Maksimov, Vaht, Murd, Harro and Bachmann2015b). Considering that there are a multitude of genes with effects expressed also in cholinergic, noradrenergic, glutamatergic, and other functions (obviously impacting perceptual behavior) and that already any optional relatively small observers’ sample can invoke covert neurobiological variability with even more varying perceptual effects, any model aspiring to be general cannot be meaningfully established. Therefore, the recommendation by R&D to test more specific models with specific priors and individual limitations (as the sources that corrupt decisions) seems right. As the number of free parameters in genetic variability is very large, the task of distilling and grouping individualized endophenotypal priors as independent experimental variables is not quick and easy. But it may have importance considering that many genes implicated in perceptual behavior are also the so-called risk genes in terms of their relation to health-related vulnerability.
R&D argue that dissociations and low degree of correspondence between confidence and accuracy make general statements about the optimality or suboptimality of perceptual decisions meaningless; similar problems are related to perceptual biases, illusions, and improbabilities (sect. 3.5.2 and 3.8.1). However, the concepts of noise and bias are ambiguous when we move from general information-theoretic perceptual decision models to real psychophysiological processes and mechanisms and want to use these models in consciousness research. For example, associative learning–based expectations tend to produce hallucination-like illusions such that observers experience auditory or visual stimuli that are objectively absent (Aru & Bachmann Reference Aru and Bachmann2017; Mack et al. Reference Mack, Erol, Clarke and Bert2016; Powers et al. Reference Powers, Mathys and Corlett2017). Importantly, the proneness to such illusions is individually variable. Bayesian prior related top-down effects can produce perceptual illusions or hallucinations that are technically noise (and must be included in noise distribution), but they are at the same time phenomenal perceptual experiences taken as veridical by the observers. Hence, this “technical noise” becomes an important factor of real awareness based behavior. On the other hand, noise can be interpreted traditionally as the factor countering sensitivity simply by decreasing the signal-to-noise ratio without any illusory non-veridical experiences. Similarly, hallucinatory experiences are technically indications of bias (type 1 bias) but are at the same time factors determining behavioral choices. The other version of the bias effect relates to its impact on decision without illusory experiences of the imperative stimuli being involved (type 2 bias). In the case of illusory non-veridical percepts, high confidence ratings are expected, but in the case of no illusory experience (but low signal-to-noise ratio), low confidence ratings are expected. Both the concept of noise and the concept of bias in the traditional perceptual decision models used for advocating optimality are underspecified or even misleading in the light of experimental data on consciousness contents and qualia of perception, including illusions and hallucinations. Behaviorally incorrect perceptual responses can be subjectively right, leading to high confidence.
With regard to the domain of decision rules, R&D note that observers may not consider all information relevant to the decision. Moreover, in most of the stimulation that is present for evaluation by the observer, there are several cues on which the decision can be based. When perceptual decision models (and also methods of scaling subjective experiences such as the Perceptual Awareness Scale [PAS]; Sandberg et al. Reference Sandberg, Timmermans, Overgaard and Cleeremans2010) use too general, underspecified evaluation criteria such as level of confidence or subjective clarity, difficulties arise in interpreting the results and explaining the underlying mechanisms and processes. The same behavioral outcome (e.g., successful discrimination or recognition) can be founded on different aspects of subjective contents of perception as different criterion contents (Jannati & Di Lollo Reference Jannati and Di Lollo2012; Kahneman Reference Kahneman1968). For behaviorist models of perceptual decisions, this may not be a problem, but for models to be used for consciousness studies, the excess generality of scales and criteria is a shortcoming. The cure would be development and use of specific rating scales for different subjective aspects of the same stimulation. This would be useful also for accounting for the effects of cue dissociation, which is an alternative source of ambiguities in addition to redundancy effects from cue combination.
To conclude, I tend to agree with R&D in that general statements about the optimality or suboptimality of perceptual decisions are meaningless unless detailed enough observer models will be put together.
Acknowledgment
I acknowledge support from IUT20-40 (Estonian Science Agency).