Early roots of information-seeking: Infants predict and generalize the value of information

Reviewer #1 (Public Review):

This study investigates the underlying mechanisms of information-seeking in infancy. Eight-month-old Dutch infants were tested in a screen-based eye-tracking task in which one of two geometrical shape cues (differing in their shape and motion) either announced the location of an upcoming reward cartoon (informative) or not (non-informative). The authors measured the infants' pupil size before the cartoon appeared. Infants showed smaller pupil sizes when presented with the informative cue as compared to the noninformative cue. The decrease in pupil size in the informative condition emerged over the course of trials whereas infants' pupil size remained unchanged in the noninformative condition. The authors interpret their findings as supportive evidence of statistical learning and generalization processes organizing infants' information-seeking.

It was a pleasure to read the paper and I think the study makes a valuable contribution to our understanding of information-seeking in infancy. The manuscript is very well written and the study is cleverly designed. My following comments are based on my reading of the manuscript and the supplemental materials. It should be noted that evaluating the details of the statistical procedure the authors used lies outside my expertise. The same applies to some decisions of the authors related to pre-processing and filtering the pupil data. I very much appreciate that the authors shared all their raw data and analysis scripts openly accessible on the Open Science Framework. The study was unfortunately not preregistered, making it difficult to trace when in the study process certain decisions or assumptions were made.

My two main concerns relate to the conceptualization and definition of information-seeking and the proposed speed of the mechanisms explaining infants' behavior. I outline my general comments below before listing some more concrete issues.

1. While reading the manuscript, I was sometimes confused about what the authors refer to when talking about information-seeking - both in terms of the broader conceptualization of the phenomenon as well as when referring to their own study. What information are infants seeking? The informative value of the cue shape in terms of their motion (because it carries information about the location of a rewarding animation)? Or is the target (the rewarding video) the information being sought? From how the study is set up, I assume the authors refer mainly to the first aspect, but I think the manuscript would benefit from some clearer distinctions and definitions of terms.

More specifically, I think it could help if the authors would specify the different aspects involved in information-seeking in the introduction (e.g., seeking information "directly", seeking cues guiding them towards information, etc.). Secondly, it would help if they would sharpen their (already in some parts existing) definitions for their study and then keep consistent with their definitions throughout the methods, results, and discussion. Is the cue the information being sought or the "behavior" (motion) of the cue? Or is the target animation the information being sought and guided via the cueing?

1. Speed of the generalization process:
   From my understanding of the study design, the shape of the geometrical shape gains informative value over time (serving as an informative cue) and the *motion* of the shape is the actual informative or non-informative visual cue in that it either reliably highlights the actual target region (or all regions). In the generalization trials, only the shape was manipulated while the motion aspect remained consistent with the previous trials. Based on infants' behavior across learning and generalization trials, the authors make an argument about two distinct processes taking place: a slower allowing to learn where to find info and a faster generalization process. Apologies if I missed something, but given that the motion remains consistent, it's maybe not surprising that the generalization trials are "faster"? Maybe the generalization process would have been slower if not only the shape had changed but if also a novel informative motion had been introduced. Also, it would be helpful if the authors could clarify what they mean by the statistical learning process being more "data-hungry" (line 274).

2. I would find it very helpful if the authors would discuss statistical learning and information-seeking processes from other possible mechanisms such as reward learning mechanisms. For example, the authors use a "rewarding" (not informative) stimulus as the target-wouldn't it be possible that the results can be also explained by reinforcement learning processes? Relatedly, in line 396 they write that they used TD learning to predict whether "information will be delivered" and contrast this with the approach being used to predict whether a reward will be delivered. But in their study reward was being delivered, too (in the form of the target), in addition to the informative motion of the cue.

Reviewer #2 (Public Review):

Summary
The study used eye tracking with a focus on pupillometry to examine how infants can learn to distinguish between informative and uninformative visual cues. Infants (n = 30, mean age = 8.2-months-old) viewed displays consisting of a sequence of stimuli: a fixation point, a central cue that predicted a subsequent informative or uninformative signal, the signal itself, and the target event (a cartoon animal, referred to as the reward). The key results are that: (1) pupil size differs depending on whether the infants anticipated an informative or uninformative signal, (2) this difference develops across trials, consistent with a slow learning process, and (3) there is rapid generalization when new shapes were introduced that shared features with the informative vs uninformative cues. The study complements a rich literature, including from this same group, showing that children are sensitive to information gains, and is interesting and important in revealing that pupil size is a physiological marker of information anticipation. We have several comments and concerns and believe that addressing them would substantially strengthen the manuscript.

Major points are related to interpretation, statistical robustness, and clarity

1. There is a tendency to overinterpret the findings.
a. Throughout, the authors interpret the findings as meaning that pupil size tracks the "value" of information; however, the results do not demonstrate conclusively whether, or what kind of value information has in this task. A natural hypothesis is that infants are intrinsically motivated to predict - i.e., value the ability to predict the target event as early as possible. In a supplementary figure, the authors present evidence that infants indeed fixate on the target event sooner after seeing informative vs uninformative cues, consistent with the idea that they use the information for improving predictions. However, those results are not fully convincing, as we detail in point 2. Most importantly, the analysis is not integrated or even mentioned in the main analyses analysis. Making the link between the pupil reaction and the use of the information would greatly strengthen the paper (whether or not the supplementary findings hold up to more thorough scrutiny). Either this link should be made and discussed, or the authors should soften their conclusions about the utility of the informative cues.

b. On line 236, the text states that the evidence "...supports the growing body of evidence indicating that infants are proactive in shaping their learning environment by searching for and focusing on information-rich stimuli". The results do not show that the infants search for information, only that they have a pupil reaction that differentiates between informative and uninformative stimuli.

c. On lines 248-249, it seems a stretch to relate the changes in pupil dilation to a shift in information value onto the cue. Without some other measure (e.g., EEG), this remains speculative. While I believe the suggestion is plausible, the language should be softened to highlight this as a follow-up research question that the present research cannot directly speak to.

2. Several findings are statistically weak and several analyses are insufficiently controlled.

a. The analysis in Supplementary Figure 2, which shows that the latencies of target fixations are shorter after informative vs uninformative cues, raises several questions.
i. We were unable to fully test these analyses as the OSF project seems to only contain latency data for 33 participants (including 22 of the 30 that remain in the final sample).
ii. The results are described as revealing a significant difference, but the 89% confidence interval of the difference contains 0. How did the authors establish significance here?
iii. How do the authors distinguish incidental fixations (which just happened to land near the target) from true predictive gaze shifts? Fixations were pooled if they occurred from 1.25 seconds before to 1 second after target onset. This is sufficient time for the eye to move in and out of the window several times. The authors should analyse the distributions of fixation durations to rule out various artifacts unrelated to target prediction.
iv. Latencies to fixation were standardized, bringing the mean across each participant to 0, and yet the statistical model includes a random intercept; is there a justification for this?
v. Standardizing removes information about whether fixations were proactive or reactive. It would be very interesting to see if/how information affects these two differently.
vi. Since informativeness was learned across trials, it seems desirable that the model should include as random effects a trial number and an interaction between trial number and informativeness. This would allow a comparison between learning to predict and the pupil reaction. Are infants who have a stronger (or earlier) pupil reaction also more likely to show stronger learning to anticipate?

b. The main finding that pupil size differs between informative and uninformative cues is based on a 3-second analysis window. This long window most likely spans many saccades, which can affect pupil size on its own or by bringing the eye on or off visual stimuli. There is no analysis to show that the statistics of saccades or fixation locations are equivalent between the two trial types - but this is necessary to convincingly rule out a spurious artifact.

c. The second main finding that the effect of informativeness grows across trials seems statistically weak. The text (line 138) states that the interaction had a beta of 0.002, which was equal to the lower border of the 89%HDI ([0.002, 0.003]). For the second claim that pupil size decreased across informative trials, the beta is -0.002, and 89% HID is non-existent - i.e., [-0.002, -0.002]. (In general, the authors should check their numbers more carefully and make sure they are presented with a degree of precision that allows the reader to interpret them meaningfully.

d. The analyses do not indicate how well the TD model fits; we are shown only that it fits better than a linear model. On line 177 a correlation analysis is mentioned between the data and model, but the statistic cited for this test on line 179 is a mean beta coefficient, so it is impossible to know what this means. An analysis of goodness of fit or, at the very least, a figure superimposing the model and data, would be much more convincing.

3. The descriptions are very unclear in some key parts of the paper

a. The description of the TD model applied to pupil learning (starting on line 391) is very unclear. The model has to include some measure of informativeness - i.e., the match between the cued and true target location - but it is unclear how this was formalized. It is also very unclear how time within the trial is incorporated (the meaning of the TDE equation).

b. The description of the generalization analysis (Fig. 5) is also very unclear. Every single sentence in it evoked some confusion, so I will go through them one by one. "A Bayesian additive model showed that infants' pupil dilation was reduced for novel cues." Reduced relative to what? "This was specific to those novel cues that shared the features of the familiar informative cues (estimated mean difference = -0.05, 89%HDI = [-0.062, -0.038])." All the novel cues shared features with the informative cues; do the authors mean the novel cues that had the critical feature indicative of the informative cue? "The size of this effect approximated the difference between conditions that were observed for familiar stimuli (estimated mean difference = -0.067, 89% HDI = [-201 0.077, -0.057])." What is "this effect"? "Crucially, this difference was not observable at the start of the task, when the familiar stimuli were first introduced (estimated mean difference = -0.007, 89%HDI = [-0.015, 0.001])." At the start of the task, the stimuli were novel, and not familiar.

Reviewer #3 (Public Review):

Summary:
The study attempts to shed light on the mechanisms underlying information-seeking in infants by investigating whether infants distinguish between informative and uninformative stimuli to resourcefully allocate their attention. The authors show that 8-month-old infants can learn whether a visual stimulus is informative or uninformative about the location of a later appearing rewarding stimulus by employing statistical regularities from the input. Specifically, infants showed decreased pupil dilation for informative over uninformative cues, which developed over the course of trials as more and more information was gathered from the input. The pattern of learning was in line with a reinforcement learning model which employed a steep learning curve in the beginning followed by a more shallow but steady learning growth over trials. After 17 trials, the authors presented novel cues that shared certain visual features with the previous stimuli and showed that pupil dilation was reduced for novel cues that shared features with the previous informative stimuli, suggesting that infants were able to generalize their acquired knowledge about the informativeness of certain features to novel stimuli. The present study adds to the existing literature about the underlying mechanisms of learning by showing that infants cannot only predict an upcoming stimulus based on statistical regularities of a preceding cue but also the informativeness of the cue itself.

Strengths:
The authors use a suitable method to test the highly relevant question of whether and how infants infer the informativeness of stimuli from experience and whether they can generalize this knowledge to new stimuli. Their experiment is carefully designed and well controlled with conditions closely matched (e.g., the shape and color of objects and the structure of each trial). Their measure of interest (i.e., pupil dilation) is also examined at a time point in each trial when the conditions are the most similar, which further points to a thought-through and careful design. This empirical data is backed up with a computational approach (using a Bayesian model and training a reinforcement learning algorithm) to elucidate the learning mechanisms at play. This approach is explained concisely to readers not familiar with the models.

The results are convincing showing a clear difference between informative and uninformative condition and development over trials. Specifically, this difference is not apparent in the first trial (Fig. 2c) but develops over time which supports a learning trajectory. The data support the authors' conclusion that infants learn about the informativeness of the object cue from the input, and the employed learning algorithms give further insights into the learning trajectory of the infants. Overall, the statistical analyses seem solid and the priors for the Bayesian models are well reported.

Data and scripts are openly available fostering transparency.

Overall, the manuscript is very well and concisely written.

Weaknesses:
The authors' conclusion that infants can generalize the acquired knowledge to similar but novel stimuli is weakened by methodological concerns regarding the analysis. It is not fully clear which trials the authors excluded and analyzed as they do not consistently report the trials in the manuscript (e.g., it is stated that after trial 17 the first generalization trial started, but also that trial 17 was excluded as the first trial of the generalization phase). As there are only a few novel trials and novel and familiar trials alternated, the inclusion or exclusion of trial analyses might have a significant impact on the results. Thus, this needs further clarification. The authors also mentioned that the novel stimuli shared relevant as well as irrelevant features, but it was not clear to me whether the authors could establish that only the relevant features contributed to the observed generalization effect.

Some methodological decisions were not explained and need justification, in particular, as the study is not preregistered. This includes, for example, the exclusion criteria and the choice not to analyze all generalization trials. Further, the authors did not perform model comparison (e.g., their model against a null model) and therefore do not report the strength of evidence for a difference in conditions.

Another weakness is that the sample sizes of 30 infants for the initial part and 19 infants for the generalization part of the experiment are rather small (especially with regard to the chosen weakly informative priors).