Illusions, Delusions, and Your Backwards Bayesian Brain: A Biased Visual Perspective. Born R.T., Bencomo G.M. Brain Behav Evol, Mar 2021. https://doi.org/10.1159/000514859
Abstract: The retinal image is insufficient for determining what is “out there,” because many different real-world geometries could produce any given retinal image. Thus, the visual system must infer which external cause is most likely, given both the sensory data and prior knowledge that is either innate or learned via interactions with the environment. We will describe a general framework of “hierarchical Bayesian inference” that we and others have used to explore the role of cortico-cortical feedback in the visual system, and we will further argue that this approach to “seeing” makes our visual systems prone to perceptual errors in a variety of different ways. In this deliberately provocative and biased perspective, we argue that the neuromodulator, dopamine, may be a crucial link between neural circuits performing Bayesian inference and the perceptual idiosyncrasies of people with schizophrenia.
Keywords: Cerebral cortexDopamineNeuromodulatorsSchizophreniaSensory systemsVision
Closing Remarks and Future Directions
This has been an admittedly biased review of several different bodies of literature, perhaps illustrating the pitfalls inherent in overly strong priors. Our aim from the start was to be provocative and, whether the ideas presented here are absolutely correct in their detail is less important than the dialogue and future studies that we are hoping to inspire. Besides, it takes only one additional inhibitory interneuron intercalated into a circuit to completely invert the sign of a predicted effect or influence! And, as noted above, simply changing the subtype of a neuromodulator’s receptor can lead to very different effects on the same circuit.
In this spirit, we close with a few thoughts on several specific areas that we think merit deeper investigation. First, at the circuit level, the mechanisms by which top-down information interacts with local circuits remain largely unknown, exacerbated by the fact that many of these interactions take place in layer 1. While modern approaches using serial-section electron microscopy (EM) have begun to flesh out the details of local circuits [Morgan and Lichtman, 2013], layer 1 has not been amenable to traditional EM-based connectomics, because, as previously noted, the vast majority of the inputs are from distant sources. However, such distant sources might soon be identifiable in serial EM reconstructions by using recently developed methods that allow neural tracing with viral vectors carrying different genetically encoded labels that are distinguishable with EM [Cruz-Lopez et al., 2018; Zhang et al., 2019]. New, nondestructive imaging methods also promise to extend the distances over which circuits can be reconstructed at the ultrastructural level [Kuan et al., 2020].
Second, most studies on the influence of top-down information on perception and cognition have been done in humans and NHPs, where tools to study circuit mechanisms are lagging compared to those in rodent models. In the future, this border zone needs to be more thoroughly investigated, both by improving our toolkit for circuit-level manipulations in NHP [Dai et al., 2015; El-Shamayleh et al., 2016; Galvan et al., 2017] and seeking out meaningful touchpoints between studies on NHPs and rodents, in the spirit of Figure 4.
Third, the mechanisms by which neuromodulators influence specific cortical circuits are poorly understood; a myriad of cellular and synaptic effects have been described but understanding the overall effects will require sophisticated computational models [Seamans and Yang, 2004].
Fourth, how circuit-level influences of neuromodulators lead to changes in perception and behavior remains deeply mysterious. This is true, not only for dopamine, but other neuromodulators as well. Chief among those that seem ripe for investigation is serotonin (5-HT), given the powerful perceptual distortions that are produced by hallucinogenic drugs, most of which are believed to act through 5-HT2A receptors [Nichols, 2004; González-Maeso et al., 2007; Halberstadt, 2015]. The historical events [Pollan, 2019] that led to these drugs being classified as “schedule 1” made them virtually inaccessible to the scientific community for many years. Thankfully, this historical influence appears to be on the wane, and we hope that perceptual scientists will make use of this powerful set of tools for future studies on perception.
Finally, the body of literature showing a reduced susceptibility to contextual visual illusions and abnormal corollary discharge in patients with schizophrenia, while suggestive, remains difficult to interpret for a variety of reasons including the fact that most of these patients are on a variety of psychoactive medications, are often condemned by their illness to extremely difficult socioeconomic situations, and frequently have other neuropsychiatric diagnoses. In this regard, several studies showing diminished top-down perceptual effects in the normal population that correlate with “cognitive-perceptual schizotypal traits” [Teufel et al., 2010; Bressan and Kramer, 2013] seem particularly promising, particularly given the possibility of conducting large-scale psychophysical studies online, using tools such as Amazon’s “Mechanical Turk” [Rajalingham et al., 2015; de Leeuw and Motz, 2016].
There is a tremendous gap between the conceptual simplicity of Bayesian inference and our understanding of the neural mechanisms that might implement it. Even such seemingly basic questions as how neural systems represent probability remain unsettled [Beck et al., 2008; Ma and Jazayeri, 2014; Haefner et al., 2016; Walker et al., 2020]. The situation might seem hopeless. Connectomics has revealed seemingly Byzantine cortical circuitry [Bock et al., 2011] which can adopt a variety of different functional modes under the influence of multiple systems of neuromodulators [Bargmann and Marder, 2013], each having scores of effects at different levels of the circuit [Seamans and Yang, 2004; Tritsch and Sabatini, 2012]. While new experimental tools to probe circuit function are surely part of the solution, ultimately, what is most needed are synthetic computational models, i.e., models that themselves represent the consensus of an entire modeling community [Bower, 2015], which can integrate results across different levels of investigation into (hopefully) simpler explanations at the level of circuit motifs that perform canonical computations [Douglas and Martin, 2007; Kouh and Poggio, 2008; Carandini and Heeger, 2011; Miller, 2016] in the service of behavioral goals [Krakauer et al., 2017].
Posts
