A stimulus-computable rational model of habituation in infants and adults

Reviewer #1 (Public review):

Summary:

This paper proposes a new model of perceptual habituation and tests it over two experiments with both infants and adults. The model combines a neural network for visual processing with a Bayesian rational model for attention (i.e., looking time) allocation. This Bayesian framework allows the authors to measure elegantly diverse factors that might drive attention, such as expected information gain, current information gain, and surprise. The model is then fitted to infant and adult participants' data over two experiments, which systematically vary the amount of habituation trials (Experiment 1) and the type of dishabituation stimulus (familiarity, pose, number, identity, and animacy). Results show that a model based on (expected) information gain performs better than a model based on surprise. Additionally, while novelty preference is observed when exposure to familiar stimuli is elevated, no familiarity preference is observed when exposure to familiar stimuli is low or intermediate, which is in contrast with past work.

Strengths:

There are three key strengths of this work:

(1) It integrates a neural network model with a Bayesian rational learner, thus bridging the gap between two fields that have often been disconnected. This is rarely seen in the cognitive science field, but the advantages are very clear from this paper: It is possible to have computational models that not only process visual information, but also actively explore the environment based on overarching attentional processes.

(2) By varying parametrically the amount of stimulus exposure and by testing the effects of multiple novel stimulus types, this work allowed the authors to put classical theories of habituation to the test on much finer scales than previous research has done.

(3) The Bayesian model allows the authors to test what specific aspects are different in infants and adults, showing that infants display greater values for the noise parameter.

Weaknesses:

Although a familiarity preference is not found, it is possible that this is related to the nature of the stimuli and the amount of learning that they offer. While infants here are exposed to the same perceptual stimulus repeatedly, infants can also be familiarised to more complex stimuli or scenarios. Classical statistical learning studies for example expose infants to specific pseudo-words during habituation/familiarisation, and then test their preference for familiar vs novel streams of pseudo-words. The amount of learning progress in these probabilistic learning studies is greater than in perceptual studies, and familiarity preferences may thus be more likely to emerge there. For these reasons, I think it is important to frame this as a model of perceptual habituation. This would also fit well with the neural net that was used, which is processing visual stimuli rather than probabilistic structures. If statements in the discussion are limited to perceptual paradigms, they would make the arguments more compelling.

Reviewer #2 (Public review):

Summary:

This paper extends a Bayesian perception/action model of habituation behavior (RANCH) to infant-looking behavior. The authors test the model predictions against data from several groups of infants and adults tested in habituation paradigms that vary the number of familiarisation stimuli and the nature of the test stimuli. Model sampling was taken as a proxy for looking times. The predictions of the model generally resemble the empirical data collected, though there are some potentially important differences.

Strengths:

This study addresses an important question, given the fundamental nature of habituation to learning and memory. Previous explanations of infant habituation have typically not been in the form of formal models, making falsification difficult. This Bayesian model is relatively simple but also incorporates a CNN to which the actual stimulus image can be presented, which enables principled predictions about image similarity to be derived.

The paper contains data from a relatively large number of adults and infants, allowing parameter differences across age to be probed.

The data suggests that the noise prior parameter is higher in infants, suggesting one mechanism through which infant and adult habituation is different, though of course, this depends on whether there is sufficient empirical evidence that other explanations can be ruled out, which isn't clear in the manuscript currently.

Weaknesses:

There are no formal tests of the predictions of RANCH against other leading hypotheses or models of habituation. This makes it difficult to evaluate the degree to which RANCH provides an alternative account that makes distinct predictions from other accounts. I appreciate that because other theoretical descriptions haven't been instantiated in formal models this might be difficult, but some way of formalising them to enable comparison would be useful.

The justification for using the RMSEA fitting approach could also be stronger - why is this the best way to compare the predictions of the formal model to the empirical data? Are there others? As always, the main issue with formal models is determining the degree to which they just match surface features of empirical data versus providing mechanistic insights, so some discussion of the level of fit necessary for strong inference would be useful.

The difference in model predictions for identity vs number relative to the empirical data seems important but isn't given sufficient weight in terms of evaluating whether the model is or is not providing a good explanation of infant behavior. What would falsification look like in this context?

For the novel image similarity analysis, it is difficult to determine whether any differences are due to differences in the way the CNN encodes images vs in the habituation model itself - there are perhaps too many free parameters to pinpoint the nature of any disparities. Would there be another way to test the model without the CNN introducing additional unknowns?

Related to that, the model contains lots of parts - the CNN, the EIG approach, and the parameters, all of which may or may not match how the infant's brain operates. EIG is systematically compared to two other algorithms, with KL working similarly - does this then imply we can't tell the difference between an explanation based on those two mechanisms? Are there situations in which they would make distinct predictions where they could be pulled apart? Also in this section, there doesn't appear to be any formal testing of the fits, so it is hard to determine whether this is a meaningful difference. However, other parts of the model don't seem to be systematically varied, so it isn't always clear what the precise question addressed in the manuscript is (e.g. is it about the algorithm controlling learning? or just that this model in general when fitted in a certain way resembles the empirical data?)