The costs of generality

Each new article of Kate Plaisted’s group is an event and should be examined with great attention, considering their usual depth and far-seeing perspective. In the current article, Davis and Plaisted (2014) build upon Plaisted’s original Reduced Generalization Theory (RGT) to propose more selective propositions whereby reduced endogenous noise, defined as any variation in neural responses that limits detection or discrimination by reducing signal-to-noise ratio, can account for several cognitive particularities in autism. Specifically, Davis and Plaisted propose that low levels of neural noise in autism influence stimulus detection/discrimination, incite transitions between perceptual and cognitive states, and increase generalization by reducing stimulus distinctiveness. The present theory differs from its predecessor by suggesting that reduced generalization/enhanced discrimination in autism is observed only in situations where parameters vary continuously. Autistic superiority on perceptual and/or cognitive tasks should therefore be greater for tasks for which stimulus intensity is defined by continuous variables, such as during low-level signal extraction, that is, pitch, luminance, and symmetry. However, perceptual superiorities are also evident during mid-level perception, consistent with loci of enhanced neural activity in experience-dependant brain regions (Samson et al., 2012). Mid-level perception is identified with pattern construction and detection, which is by nature a non-linear, threshold-type phenomenon which may narrow the applicability of the present theory to autistic cognition. For example, when the authors invoke superior distinctiveness and clarity of local elements to explain local bias in Navon-type stimuli, it is not clear what the added value of noise to that of enhanced pattern detection, a process that can be deduced from this group’s early articles concerning visual search performance. Also, as reduced noise is argued to encourage transitions between neural states, reduced noise would explain cognitive inflexibility, by trapping attention or perception in stable states. This use of the “attractor metaphor,” albeit inherently polysemic due to its mathematical nature, goes far beyond the disappointing executive account of autistic cognitive rigidity. It fits nicely with the longer visual inspections and more generally interest for perceptual characteristics of objects and suggests that accuracy and rigidity are intrinsically linked. This should result in novel experiments, with testable predictions, and is plausibly one of the more interesting aspects of their theory. Finally, it is argued that noise increases generalization and reduces stimulus distinctiveness; conversely, reduced noise should account for limited or slow categorization. Whereas this idea remains appealing, available literature indicates that autistic cognition does not follow this rule as much as one should expect. The concept of reduced generalization lacks empirical support in laboratory experiments, and particularly in front of intact implicit learning in autism (Foti et al., 2014), contrary to what autistic behavior in natural settings seems to suggest.

Using the author’s own words, noise is a malleable concept. It is a construct originating from information theory and does not refer to a concretely delineable cognitive loci or apparatus implemented in the human brain. Despite being presented as noise “in neural networks,” the tenets of this theory need to be implemented and grounded within mechanistic and actual microstructural constructs. But will it help us find new mechanistic explanations, or only a posteriori allow regrouping these explanations in a meaningful category?

Balancing between description and explanation, the search for an economic—and therefore, general, multilevel explanation risks producing principles open to a multitude of possible implementations. Noise is as far from what it tries to explain as Frith and Happé’s (1994) Central Coherence, Pellicano and Burr’s (2012) hypo priors, and, but possibly to a lesser extent, Van de Cruys et al.’s (in press) predictive coding theory. We agree that to explain autistic perceptual, mnemonic, and reasoning superiorities, one may have to look outside of perception—we did the same effort in extending Enhanced Perceptual Functioning to Veridical Mapping and then to Trigger-Threshold-Target (Mottron et al., 2013, 2014) models. However, to paraphrase a magnificent statement from this article, there is an inverse relationship between local and global noise, such that a decrease in one will reflect an increase in the other, which I believe is true also of theories. The cost of generalization in this case may limit the heuristic use of this concept. According to Plaisted et al. (2009) herself, one should be sometime deaf to the sirens of symmetry when modeling biology.