“Switching” between fast and slow processes is just reward-based branching

Abstract

Shortcuts to goals are rewarded by faster attainment and punished by more frequent failure, so selection of the various kinds – heuristics, cached sequences (habits or macros), gut instincts – depends on reward history just like other kinds of choice. The speeds of shortcuts lie on continua along with speeds of deliberation, and these continua have no obvious separation points.

This target article (TA) follows on De Neys's recent proposal “that trying to answer the core single vs dual process model debate is pointless for empirical scientists” so it is “time to move on” (De Neys, 2021, Introduction). What he proposes in the TA is “a more viable dual process architecture” (target article Abstract), which is “orthogonal [to] whether the difference between the two types of processing should be conceived as merely quantitative or qualitative.” Nevertheless, he argues for two qualitatively different processes, perhaps characterized by the 14 different properties he listed in the earlier article (fast, effortless, affective, automatic… vs. slow, effortful, affectless, controlled…; De Neys, 2021, p. 4), which he calls here simply fast and slow. He demonstrates the flaws in dual-process theories' usual assumptions: that the two processes must operate separately (“exclusivity”), and that there must be a “switch feature… by which a reasoner can decide to shift between more intuitive and deliberate processing”; but he pulls back from other authors' proposal that “we simply abandon the dual process enterprise.” His refutation of the authors who have favored a single, quantitatively based decision process is just to point out that “some responses require more deliberation than others,” which would not seem to require a dichotomy.

Dual-process models have admittedly been popular over the years, beginning with Plato's wild versus well-behaved chariot horses. In addition to De Neys's fast versus slow examples, choice making has been described as passionate versus reasonable, impulsive versus reflective, myopic versus far-sighted, hot versus cool, and model-free versus model-based, among others. De Neys also includes as fast the products of “automatization,” by which repeated sequences of choices “will be elicited intuitively.” In addition, brain imaging has found evidence for steep-discounting versus shallow-discounting brain centers (McClure et al., 2004; van den Bos & McClure, 2013).

However, as De Neys himself concludes, there is no operation that system 2 can perform that system 1 cannot, and “thinking always involves a continuous interaction between system 1 and system 2 application” (target article, sect. 3.4, para. 5). Other authors have pointed out obvious problems with the dual approach: If there are distinct systems, there must be more than two of them, because the properties attributed to the two systems do not reliably occur together (Zbrodoff & Logan, 1986); in particular, automatic processes may or may not be affectively arousing (Ainslie, 2021). Furthermore, the listed properties such as effort, affect, and speed itself are themselves continua. If the two whole lists of properties really constitute discrete systems, there should be natural breaks in the continua from fast to slow, and the breaks should occur at equivalent levels in the n dimensions. As a negative example, the only obvious break in transparency would be too-fast-to-introspect versus not-too-fast-to-introspect, which would not define different kinds.

Most “type 1” processing in humans comprises sequences that have been automatized, macros (or habits) that call up other macros. In language, a squiggly line is interpreted as a letter, a sequence of letters is interpreted as a word, a series of words forms a concept (or cliché). All highly automatized; but if I was to find an anomaly – no, it should be “were to find an anomaly” – my ear would be quick to re-set it. This should not require a distinct system. Even if I stopped to ponder the use of the subjunctive, I would just be trying out sequences I had previously automatized. Likewise, as my calculation proceeds from 2 + 2 through, say 8 + 8, to 64 + 64, and so forth, at some points my mind will pause to find component automatizations; but is there a point where the pause divides two systems?

The strongest case for separate processes might be based on the activities of separate sites in the brain, but even here true separation is doubtful. The dorsolateral striatum (putamen) is differentially active when repeated connections have been cached to form macros, whereas the dorsomedial striatum (caudate) is more active during flexible behavior; but their functioning has been observed to be integrally combined (Dolan & Dayan, 2013; Keramati et al., 2016). Similarly, the existence of separate steep and shallow reward discount centers in the brain is controversial (Kable & Glimcher, 2007; Lempert et al., 2019). If there do exists anatomically separate response-selection systems in the brain, the best candidates would be those for motivational salience and (supposedly separate) reward, governing the attraction of attention and behavioral approach/avoidance, respectively (Berridge & Robinson, 1998). But even here, salience and behavior selection are correlated with activity in mostly the same brain regions (Kim et al., 2021); and when even threatening stimuli are voluntarily gated out, attention to them must have been weighed in the common marketplace of reward (see Ainslie, 2009).

The professed scope of the TA's model is universal, but except for its reference to cupcakes its examples are cognitive searches for correct solutions to puzzles, rather than choices among competing rewards. Accordingly, “the peak activation strength of an intuition reflects how automatized or instantiated the underlying knowledge structures are (i.e., how strongly it is tied to its eliciting stimulus).” This rather Pavlovian convention hampers the model's application to goal-directed activities. By contrast, it is feasible to model the selection of all learnable processes which can replace each other using the amount and timing of their contingent reward (Ainslie, 2017). The sources of reward – consumption goods, ethical goods, social cues, puzzle solutions, signal detections, emotions, the satisfaction of urges – as well as their speeds of onset, are miscellaneous. It should not matter that some of their subroutines involve particular parts of the brain (for instance the amygdala – Aqino et al., 2020 – or hippocampus – Gauthier & Tank, 2018), as long as their weights are ultimately comparable to each other. Likewise, the weighing process may or may not involve a specific site, such as the orbitofrontal cortex (Bartra et al., 2013; Levy & Glimcher, 2012), a set of interacting sites (Krönke et al., 2020), or no identifiable dwelling place (Dohmatob, Dumas, & Bzdok, 2020). In reward research, the adoption of millisecond-specific electroencephalography (for instance, Sambrook et al., 2018) promises to give precise evidence about branching to fast, slow, and intermediate processes.