How does curriculum learning influence cognitive map formation?

Reviewer #1 (Public review):

This paper investigates how different learning curricula influence the way that humans piece together directly experienced transitions into a broader cognitive map. When adjacent learning trials were grouped within rows or columns of the map, subsequent navigation through the map was weaker than when adjacent learning trials came from disjoint spaces in the map. The authors speculate that the grouped curriculum resulted in mental fragmentation that made navigation across space more difficult later on.

This is an interesting paradigm that contributes useful new findings in the domain of map learning to the growing literature on curriculum learning. The evidence for a difference between conditions is highly compelling, but, as the authors are very transparent in acknowledging in the Discussion, the evidence for their proposed mechanism - mental fragmentation under grouped learning - is somewhat weak. The study thus presents an intriguing empirical result but not an ironclad mechanistic account.

An alternative - by their account, "less interesting" - explanation is that grouped learning was easier because trials in close succession had overlapping elements, and so participants were not trying as hard or as engaged. There is a literature on spaced (as opposed to massed) learning being better for subsequent memory because it increases retrieval effort. It seems very plausible that this could be going on here, and the control experiment reported in the supplement would not help to rule this out. This literature deserves some discussion.

The Introduction focuses entirely on literature showing advantages in grouped over intermixed learning, setting that up as the most well-motivated expectation from the literature. Upon finding the opposite, the Discussion then mentions that interleaving has been found to be useful in "applied domains", but then returns to how surprising this is in light of recent findings in the category learning literature. But there is a substantial earlier literature on interleaved vs blocked curricula in category learning, very often finding advantages for interleaving. See, e.g., Carvalho & Goldstone, 2015, for a review. There is also a paper showing interleaving advantages in associative inference, Zhou et al., 2023, JEP:G, which is very relevant to several of the discussion section paragraphs. Thus, the treatment of the prior curriculum learning literature is currently sparse.

Reviewer #2 (Public review):

I think this paper is an excellent and timely contribution. It clearly shows that learning overlapping relationships in a disjoint training schedule (where the overlaps are not encountered close together in time) appears to aid the formation of an integrated associative memory structure (a cognitive map) and supports generalisation. I believe the methods are sound and the results are clear. I only have a couple of methodological questions that may not warrant any changes to the paper (or only very minor changes/additions):

(1) The mixed effects models did not include random slopes for the within-subject factors ("spatial manipulation" and "block"), and so the corresponding fixed effect inferences may be unsafe. Having said that, it is likely that including these slopes may not be warranted given their contribution to the model's fit. I recommend that the authors check this.

(2) The mixed effects models for accuracy appear to model average performance across trials rather than using a generalised linear model with a (e.g.) logit link function and the binomial distribution to characterise performance. I think this is a little sub-optimal, as the latter is often more sensitive. Nonetheless, it is not in any way wrong; the results are clear enough as is, and there may be a good reason to avoid a non-linear link function, which can alter the interpretation of effects close to the ceiling and floor.

I think the introduction and/or discussion would benefit from contrasting their results with Berens & Bird (2022, PLOS Comp Bio). In this paper, it is shown that blocking the training of discriminations in a linear hierarchy (what we call progressive training) substantially benefited transitive inference performance. This seems at odds with the author's finding that "participants struggle to integrate information across rows and columns, i.e. across groups of transitions that were trained separately in time".

I would really like to know what the authors think about this discrepancy (or, indeed, whether they think there is one at all). Is it possibly because "progressive" learning is some combination of "grouping", "blocking" and "chaining" (where there is a structured overlap between adjacently trained relationships)? Or is it something else, e.g., that there is a fundamental difference between learning associations and discriminations (personally, I lean on this explanation)?

Relevant to this, the authors note that their "findings do contradict recent reports from the category learning literature, where blocking seems to help learning and generalisation (Dekker et al., 2022; Flesch et al., 2018; Noh et al., 2016). It may be that where the goal is not to learn a complex knowledge structure - like a map - but simply to compress exemplars by mapping them onto a smaller number of labels - the benefits of blocking emerge." However, the benefit of progressive (blocked) training in my own work was observed in a task that required learning a complex/relational structure in the form of a transitive hierarchy, which theoretical accounts suggest depends on learning map-like representations (Whittington et al., 2020).

Reviewer #3 (Public review):

Summary:

This study examines how training regimes influence the formation of cognitive maps. Participants learned two relational maps over three days through pairwise transitions: one map was trained with grouped sequences that followed rows or columns, while the other was trained with disjoint transitions sampled randomly across the map. In addition, the study manipulated the temporal spacing of training blocks (blocked vs. semi-blocked) and tested whether the results generalized across two map geometries (a 5×5 grid and a 4×4 torus).

Furthermore, they run a follow-up experiment (or condition) testing rows and columns shuffled in the grouped condition.

While grouped training produced better performance during learning, the authors report that disjoint training led to superior performance at test on tasks probing the global map knowledge.

Summarising experimental design:

(1) Map geometry (between-subjects): 5×5 grid vs 4×4 torus

(2) Training block schedule (between-subjects): Blocked vs Semi-blocked

(3) Training regime/transition sampling (within-subject): Grouped or Disjoint (Day 1 and Day 2)

Strengths:

The study addresses a clear and timely theoretical question about how the training regime affects the formation of cognitive maps. A further strength is the well-controlled experimental design, allowing the authors to test their hypotheses in a systematic and informative way.

Weaknesses:

(1) If I understood correctly, participants learned one map on the first day and the other on the second day, with the training regime (grouped vs. disjoint) counterbalanced across maps. This raises the possibility that experience with one training regime on day one could influence performance on the second day. For example, it would be interesting to examine whether participants who experienced the disjoint regime first showed any differences when learning the grouped regime on the following day. While it may be difficult to fully disentangle such transfer effects from the main training regime effects, it would be informative to test whether performance on the second day depends on the regime experienced on the first day (e.g., whether prior exposure to the disjoint regime predicts performance on the subsequent grouped training, but not vice versa).

(2) The author mentions a control experiment. Did the participants in the control experiment complete only the training phase or also the testing tasks used in the main experiment? If testing was included, it would be informative to report whether performance at test was comparable to that observed in the main experiment. Given that this condition appears to involve blocked transitions while moving across both rows and columns, I would expect performance to fall somewhere between the grouped and disjoint conditions.

(3) Participants' performance did not differ between conditions in the map reconstruction task, suggesting that participants in both the grouped and disjoint regimes were ultimately able to form a cognitive map. Was this task always administered last during the testing session? I wonder whether the explicit request of the reconstruction task could have influenced participants' awareness of the map structure.

(4) The manuscript describes the study as consisting of four experiments (two groups per map shape, differing in the blocked versus semi-blocked schedule). However, based on the design described in the Methods, this appears more accurately characterized as a single experiment with two between factors: map geometry (grid vs. torus) and blocking schedule (blocked vs. semi-blocked) manipulated between participants, and training regime (grouped vs. disjoint) manipulated within participants.

(5) It is not entirely clear to me from the Results section whether performance at test differed between the two map geometries (grid and torus), or whether the reported effects of training regime were consistent across them.