Benefits of commitment in hierarchical inference

Abstract

Humans have the tendency to commit to a single interpretation of what has caused some observed 1 evidence rather than considering all possible alternatives. This tendency can explain various forms 2 of confirmation and reference biases. However, committing to a single high-level interpretation 3 seems short-sighted and irrational, and thus it is unclear why humans seem motivated to pursue 4 such strategy. 5 In a first step toward answering this question, we systematically quantified how this strategy affects 6 estimation accuracy at the feature level in the context of two universal hierarchical inference tasks, 7 categorical perception and causal cue combination. Using model simulations, we demonstrate 8 that although estimation is generally impaired when conditioned on only a single high-level inter-9 pretation, the impairment is not uniform across the entire feature range. On the contrary, compared 10 to a full inference strategy that considers all high-level interpretations, accuracy is actually better 11 for feature values for which the probability of an incorrect categorical/structural commitment is rel-12 atively low. That is to say, if an observer is reasonably certain about the high-level interpretation 13 of the feature, it is advantageous to condition subsequent feature inference only on that particular 14 interpretation. We also show that this benefit of commitment is substantially amplified if late noise 15 corrupts information processing (e.g., during retention in working memory). Our results suggest 16 that top-down inference strategies that solely rely on the most likely high-level interpretation can 17 be favorable and at least locally outperform a full inference strategy. 18 performance 20 25 rate description of human behavior in a broad range of tasks associated with perception (Knill and 26 Richards, 1996), cognitive reasoning (Griffiths et al., 2010), economic decision-making (Summer-27 field and Tsetsos, 2012) and motor control (Wolpert, 2007). Furthermore, mental disorders such 28 as autism and schizophrenia have been directly linked to specific computational deficiencies of this 29 inference process (Lieder et al., 2019; Jardri and Deneve, 2013). Except for simple estimation and 30 decision-making tasks (Ernst and Banks, 2002; Körding and Wolpert, 2004; Stocker and Simon-31 celli, 2006), the generative models of these tasks are hierarchical. Object recognition is a simple 32 example of a hierarchical inference task where, at the top of the hierarchy, object categories are 33 defined as specific distributions over some lower-level feature representation potentially across 34 multiple levels of feature integration. Noisy observations at the lowest feature level then allow 35 to infer the corresponding object category by inverting the hierarchical generative model (Ker-36 sten et al., 2004). Various studies have shown that human behavior in such tasks is accurately 37 described by Bayesian statistics that fully integrate all information in the hierarchical generative 38 model (i.e., full...

Belief states and categorical-choice biases determine reward-based learning under perceptual uncertainty

“…Therefore, the belief state partly discarded decision-incongruent information in favor of self-consistency across the task stages (Peters et al, 2017). As argued in both Peters et al (2017) and Luu and Stocker (2018), these seemingly sub-optimal behaviors might be resource-rational in natural settings with many hidden states and under limited cognitive capacities (see also Qiu et al (2020)).…”

Subjective confidence influences word learning in a cross-situational statistical learning task

“…In the current cross-situational learning task, where each trial provides new evidence that may help to refine the word-meaning mapping, a bias to persist in one's own belief and to discard new information may seem to be maladaptive. However, in real word learning scenarios, when one's own lexical knowledge is thought to be reliable, it may be rational to ignore conflicting information, presumably outliers or uninformative observations, to increase the robustness of the learning process, especially as the hypothesis space of possible meanings is vast and not explicit (see relatedly Oaksford & Chater, 1994;Qiu, Luu, & Stocker, 2020;Tsetsos et al, 2016). Perhaps most obviously, it is hard to see how inferential processes such as mutual exclusivity would work without assessing the reliability of one's knowledge about word meanings: either one would apply mutual exclusivity immediately, using words for which one does not have much evidence to learn the meaning of other words (which would lead to a cascade of errors), or one would never apply it because it is objectively impossible to eliminate all uncertainty.…”

Point estimate observers: A new class of models for perceptual decision making

“…Zamboni, Ledgeway, McGraw, & Schluppeck, 2016). Such a scheme is suboptimal (Fleming, Maloney, & Daw, 2013), although it can be beneficial in the presence of later distortions of the stimulus representation (Qiu, Luu, and Stocker 2020;and closely related Li, Castañón, Solomon, Vandormael, and Summerfield 2017). Also in contrast to our point estimate observers, the original categorisation decision is based on full Bayesian inference.…”