A key move in Firestone & Scholl's (F&S's) impressive, skeptical review of putative top-down effects on perception is to bracket many attentional effects as “peripheral” – as working by altering the inputs to a cognitive process rather than by altering the processing itself. The deflationary thought (developed in sect. 4.5 of the target article) is that many attentional effects may be rather like turning one's head towards some interesting stimulus. Head-turning may indeed be driven top-down in ways sensitive to what the agent knows and intends. But in such cases, it seems that knowledge and intentions simply alter (via action) what the visual system gets as input. This idea is fully compatible with subsequent visual processing occurring in an inflexible, encapsulated (knowledge-impermeable) manner. To show genuine (interesting, nontrivial) top-down effects upon perception therefore requires showing that such effects impact not just the inputs but the processing itself, and do so in ways sensitive to what the agent (or better, I'll argue, the system) knows. F&S (sect. 5, para. 2–3) appeal to this possibility to “deflate” empirical evidence (such as Carrasco et al. Reference Carrasco, Ling and Read2004) suggesting that selective attention alters what we see even in cases where no overt action (such as head-turning or visual saccade) is involved. Selective attention, the authors concede, may indeed enhance specific aspects of the visual input. But such effects may be functionally similar to turning one's head, altering inputs that are then processed using a modular, inflexible system.
It is certainly conceivable that attention might always or mostly operate in just the blunt fashion that F&S suggest. But recent work on the so-called Bayesian brain, and (especially) related neurocomputational proposals involving predictive processing, already suggest a different, and much richer, mode of influence. That work (see Bastos et al. Reference Bastos, Usrey, Adams, Mangun, Fries and Friston2012, and reviews in Hohwy Reference Hohwy2013; Clark Reference Clark2013; Reference Clark2016) depicts online perception as subject to two distinct but interacting forms of top-down influence. The first involves simple prediction: Downwards and lateral connections are said to be in the business of predicting the ongoing stream of sensory stimulation, using a generative model that aims to construct the inputs using stored knowledge. That constructive process is answerable to the evidence in the sensory stream, and mismatches yield prediction error signals that drive further processing. But that answerability is itself subject to another form of knowledge-based top-down influence. This second mode of influence involves the so-called precision-weighting of select sensory prediction error signals. Precision-weighting mechanisms alter the influence (postsynaptic gain) of specific prediction error signals so as to optimize the relative influence of top-down predictions against incoming sensory evidence, according to changing (mostly subpersonal) estimations of the context-varying reliability of the sensory evidence itself (see Feldman & Friston Reference Feldman and Friston2010).
This suggests a far richer model of attentional effects than the one suggested by F&S. Attention, thus implemented, is able to alter the balance between top-down prediction and bottom-up sensory evidence at every stage and level of processing. That same process enables such systems to reconfigure their own effective architectures on a moment-by-moment basis. They can do so because altering the precision-weighting on specific prediction error signals alters the influence that one neural area or processing regime has on other (specific) neural areas or processing routines (see, e.g., den Ouden et al. Reference den Ouden, Daunizeau, Roiser, Friston and Stephan2010). The result is a highly flexible cognitive architecture in which attention (construed as precision estimation) controls the flexibility.
In searching for possible top-down effects on perception, the authors lay great stress on the impact (or lack of impact) of top-level agentive reasons and intentions on the fine-grained course of processing. But this emphasis, from the Bayesian/Predictive Processing perspective, is also potentially misleading. In such models, the impact of top-down processing is best identified with the joint impact of priors and context-variable precision estimations. The question of how such priors and precision estimations interact with top-level agentive reasons and intentions is then a further (complex and important) one. But regardless of how that story goes, it seems likely that much of the knowledge that impacts online processing will be sub-agential, involving unconscious estimations of the reliability, in context, of various kinds of information for the task at hand.
Suppose, to take a concrete example, that I decide to look for my car keys on a crowded desktop. Suppose further that that decision automatically generates a cascade of altered estimations concerning the reliability or salience of different prediction error signals calculated as the processing unfolds, and that these alterations impact what we see as we scan the desktop. Once thus “seeded” by my top-level decision to seek out the keys, the unfolding flow of these effects could be automatic, reflecting nothing further about my intentions or goals. But wouldn't that still amount to a legitimate and interesting case of a top-down effect on perception? Such cases (memory-based variants of which have been experimentally explored by Nobre et al. Reference Nobre, Griffin and Rao2008) may look superficially like ones in which attention “merely” enhances certain inputs. But if attention alters precision estimations at many levels of processing, that begins to look much more like the kind of profile that F&S demand – a profile in which the processing itself systematically changes in response to changing top-level goals and intentions (for some early explorations in these ballparks, see Feldman & Friston Reference Feldman and Friston2010; Vossel et al. Reference Vossel, Mathys, Daunizeau, Bauer, Driver, Friston and Stephan2014 – see also Gazzaley & Nobre Reference Gazzaley and Nobre2012).
F&S (sect. 4.5) do allow that not all attentional effects need be (in their sense) “peripheral,” and that Bayesian/Predictive Processing accounts may already suggest deeper modes of influence. But to take this caveat seriously should, I think, fundamentally alter the shape of the debate. It should alter the shape of the debate because attention (conceived as the process of optimizing precision estimations) then emerges as a deep, pervasive, and entirely nonperipheral player in the construction of human experience – one that acts not as a simple spotlight but as a subtle tool capable of altering the flow and detail of online processing in multiple ways.
The practical upshot is that F&S's downgrading of most attentional effects to simple alterations in what is given as input should be treated with caution. That downgrading is inconsistent with powerful emerging neurocomputational frameworks that depict attentional effects as reaching deep into the underlying processing regime. The conceptual contrast here is stark enough. But teasing these scenarios apart will require new experimental paradigms and some delicate model-comparisons.
A key move in Firestone & Scholl's (F&S's) impressive, skeptical review of putative top-down effects on perception is to bracket many attentional effects as “peripheral” – as working by altering the inputs to a cognitive process rather than by altering the processing itself. The deflationary thought (developed in sect. 4.5 of the target article) is that many attentional effects may be rather like turning one's head towards some interesting stimulus. Head-turning may indeed be driven top-down in ways sensitive to what the agent knows and intends. But in such cases, it seems that knowledge and intentions simply alter (via action) what the visual system gets as input. This idea is fully compatible with subsequent visual processing occurring in an inflexible, encapsulated (knowledge-impermeable) manner. To show genuine (interesting, nontrivial) top-down effects upon perception therefore requires showing that such effects impact not just the inputs but the processing itself, and do so in ways sensitive to what the agent (or better, I'll argue, the system) knows. F&S (sect. 5, para. 2–3) appeal to this possibility to “deflate” empirical evidence (such as Carrasco et al. Reference Carrasco, Ling and Read2004) suggesting that selective attention alters what we see even in cases where no overt action (such as head-turning or visual saccade) is involved. Selective attention, the authors concede, may indeed enhance specific aspects of the visual input. But such effects may be functionally similar to turning one's head, altering inputs that are then processed using a modular, inflexible system.
It is certainly conceivable that attention might always or mostly operate in just the blunt fashion that F&S suggest. But recent work on the so-called Bayesian brain, and (especially) related neurocomputational proposals involving predictive processing, already suggest a different, and much richer, mode of influence. That work (see Bastos et al. Reference Bastos, Usrey, Adams, Mangun, Fries and Friston2012, and reviews in Hohwy Reference Hohwy2013; Clark Reference Clark2013; Reference Clark2016) depicts online perception as subject to two distinct but interacting forms of top-down influence. The first involves simple prediction: Downwards and lateral connections are said to be in the business of predicting the ongoing stream of sensory stimulation, using a generative model that aims to construct the inputs using stored knowledge. That constructive process is answerable to the evidence in the sensory stream, and mismatches yield prediction error signals that drive further processing. But that answerability is itself subject to another form of knowledge-based top-down influence. This second mode of influence involves the so-called precision-weighting of select sensory prediction error signals. Precision-weighting mechanisms alter the influence (postsynaptic gain) of specific prediction error signals so as to optimize the relative influence of top-down predictions against incoming sensory evidence, according to changing (mostly subpersonal) estimations of the context-varying reliability of the sensory evidence itself (see Feldman & Friston Reference Feldman and Friston2010).
This suggests a far richer model of attentional effects than the one suggested by F&S. Attention, thus implemented, is able to alter the balance between top-down prediction and bottom-up sensory evidence at every stage and level of processing. That same process enables such systems to reconfigure their own effective architectures on a moment-by-moment basis. They can do so because altering the precision-weighting on specific prediction error signals alters the influence that one neural area or processing regime has on other (specific) neural areas or processing routines (see, e.g., den Ouden et al. Reference den Ouden, Daunizeau, Roiser, Friston and Stephan2010). The result is a highly flexible cognitive architecture in which attention (construed as precision estimation) controls the flexibility.
In searching for possible top-down effects on perception, the authors lay great stress on the impact (or lack of impact) of top-level agentive reasons and intentions on the fine-grained course of processing. But this emphasis, from the Bayesian/Predictive Processing perspective, is also potentially misleading. In such models, the impact of top-down processing is best identified with the joint impact of priors and context-variable precision estimations. The question of how such priors and precision estimations interact with top-level agentive reasons and intentions is then a further (complex and important) one. But regardless of how that story goes, it seems likely that much of the knowledge that impacts online processing will be sub-agential, involving unconscious estimations of the reliability, in context, of various kinds of information for the task at hand.
Suppose, to take a concrete example, that I decide to look for my car keys on a crowded desktop. Suppose further that that decision automatically generates a cascade of altered estimations concerning the reliability or salience of different prediction error signals calculated as the processing unfolds, and that these alterations impact what we see as we scan the desktop. Once thus “seeded” by my top-level decision to seek out the keys, the unfolding flow of these effects could be automatic, reflecting nothing further about my intentions or goals. But wouldn't that still amount to a legitimate and interesting case of a top-down effect on perception? Such cases (memory-based variants of which have been experimentally explored by Nobre et al. Reference Nobre, Griffin and Rao2008) may look superficially like ones in which attention “merely” enhances certain inputs. But if attention alters precision estimations at many levels of processing, that begins to look much more like the kind of profile that F&S demand – a profile in which the processing itself systematically changes in response to changing top-level goals and intentions (for some early explorations in these ballparks, see Feldman & Friston Reference Feldman and Friston2010; Vossel et al. Reference Vossel, Mathys, Daunizeau, Bauer, Driver, Friston and Stephan2014 – see also Gazzaley & Nobre Reference Gazzaley and Nobre2012).
F&S (sect. 4.5) do allow that not all attentional effects need be (in their sense) “peripheral,” and that Bayesian/Predictive Processing accounts may already suggest deeper modes of influence. But to take this caveat seriously should, I think, fundamentally alter the shape of the debate. It should alter the shape of the debate because attention (conceived as the process of optimizing precision estimations) then emerges as a deep, pervasive, and entirely nonperipheral player in the construction of human experience – one that acts not as a simple spotlight but as a subtle tool capable of altering the flow and detail of online processing in multiple ways.
The practical upshot is that F&S's downgrading of most attentional effects to simple alterations in what is given as input should be treated with caution. That downgrading is inconsistent with powerful emerging neurocomputational frameworks that depict attentional effects as reaching deep into the underlying processing regime. The conceptual contrast here is stark enough. But teasing these scenarios apart will require new experimental paradigms and some delicate model-comparisons.