NOTE-950: Ma & Kragel — map-like emotion knowledge in hippocampus and vmPFC

Contribution

The paper tests the proposal (its ref. 43, Ma, Vafaie and Kragel 2025) that hippocampal–prefrontal systems hold emotion concepts as a cognitive map: hippocampus binding particular concepts to locations, entorhinal cortex and vmPFC holding the relational structure. It reanalyses the public Emo-FilM fMRI dataset, decoding film-by-film emotion ratings and the activations of a Tolman–Eichenbaum Machine (TEM). After it there is a dissociation in naturalistic viewing: hippocampus carries more about emotion categories, vmPFC more about positions in a valence–arousal space, and a relational-memory model’s two layers line up with that split. Entorhinal cortex did not show the predicted structural code.

Key insight

Treat emotion concepts as places. A map-learning model that separates “what” (content, layer p) from “where” (relational structure, layer g) and binds them yields unit activations that can be read out of the hippocampus (small-scale p) and vmPFC (large-scale g) while people watch films. The inference is from decodability of a model’s units, weighted by others’ ratings, to a shared representational format. Nothing in the study watches a map being learned or used.

Assumptions

• Passive viewing, no task. 30 participants (29 analysed) watched 14 films, mean 11 min 26 s, over four sessions (p. 8). “Although our experimental paradigm was not interactive” (p. 8). They made no responses during scanning that the paper uses.
• The labels are other people’s judgements. “Emotion ratings … were collected from 44 independent participants … with no overlap between the two samples” (p. 9). Each rated six items continuously with a mouse on their own computers; consensus ratings averaged three or four raters per item and film. The same consensus time series is the decoding target for every fMRI participant. “the emotion concepts we decoded were based on self-report ratings from independent observers” (p. 7).
• Items. 13 discrete emotions; valence from “good” and “bad”, arousal from “calm” and “at ease” (p. 9). Affective space is a 2D metric MDS of 1 − r between category rating time series concatenated across films (p. 9), so its axes are derived from co-occurrence in the ratings and are not valence and arousal by definition.
• TEM. Trained for 32,000 iterations on 16 random 11 × 11 grids with 45-dimensional one-hot observations. “At each time step, it takes the current sensory observation (x) and an action (a) as inputs, and outputs a predicted sensory observation for the next timestep” (p. 9). Actions (up, down, left, right, stay) are chosen with equal probability. The “emotion environment” places each category at the grid node nearest its MDS position; the agent then walks it randomly for 12,100 steps, and per-category activations come from the second half (pp. 9–10). Film-time TEM activity is the rating-weighted average of those per-category activations.
• Decoding. PLS regression (20 components) from voxel time series to each target, leave-one-film-out, performance as the Fisher-z correlation of predicted and observed time series (p. 10). Group inference by sign-flip permutation (10,000) on mixed-effects contrasts, CIs by parametric bootstrap (p. 11).

Key results

• Hippocampus: categories over space (p. 3). Category ratings z = 0.0525 [0.0454, 0.0592] against affective-space locations z = 0.0428 [0.0347, 0.0505]; Δz = 0.0097, d = 0.4946, p < .001.
• A shared hierarchy (p. 4). Similarity among decoded category time series, leave-one-participant-out: full model Spearman r = .978, valence model r = .842, residual after valence r = .486.
• Long axis (p. 4). Posterior hippocampus decoded fine-grained categories better as granularity increased (Δz = 0.0101, d = 0.7791, p = .003). Amygdala below posterior hippocampus for categories (Δz = −0.0117, p = .001), comparable to anterior.
• vmPFC: space over categories (p. 4). Affective space z = 0.0956 [0.0831, 0.1076] against categories z = 0.0838; Δz = 0.0118, p = .020 (d printed as 0.0118). Region × target double dissociation with hippocampus Δz = 0.0215, d = 0.9566, p = .001. Entorhinal cortex and amygdala did not differ between targets.
• TEM (pp. 5–6). Hippocampus decoded layer p better than g (Δz = 0.0017 [0.0001, 0.0034], d = 0.3773, p = 0.015), more so at small scales (Δz = 0.0352, d = 1.3232, p < .001); small-scale p better in posterior than anterior hippocampus (Δz = 0.0052, p = .002). vmPFC decoded large-scale g better than p (Δz = 0.0712, d = 2.6747, p < .001), and better than hippocampus or entorhinal cortex did. Three-way region × layer × scale interaction F(8, 616) = 12.44, p < .001. Controlling for decoded p reduced the hippocampal readout of categories (Δz = 0.0042) more than of space (0.0014): p accounts for “some, but not all” of the category advantage (p. 6).

Claims

id	claim	strength	support
C1	Hippocampal patterns during film viewing predict category ratings better than affective-space positions; vmPFC the reverse	moderate	leave-one-film-out PLS, n = 29, double dissociation p = .001; small absolute correlations
C2	Decoded hippocampal category structure is graded, shared across participants and mainly organised by valence	weak to moderate	r = .978; but every participant’s decoder is trained on the same consensus targets
C3	Posterior hippocampus carries finer-grained emotion distinctions	moderate	Δz = 0.0101, p = .003; some categories reverse (p. 8)
C4	Hippocampus resembles TEM’s conjunctive layer and vmPFC its structural layer	moderate as decodability, weak as mechanism	Figs. 5, S16; one model, no rival model fitted
C5	Map-like emotion representations “could result from learning the statistical regularities of event transitions”	conjecture	from the model; no learning measured in people (p. 8)
C6	The valence–arousal map seen in self-report may be computed by hippocampal systems “rather than being an innate structure”	conjecture	Discussion, p. 8

Method

Emo-FilM fMRI (OpenNeuro ds004892) and ratings (ds004872). Masks: Julich-Brain hippocampus and entorhinal cortex, hippocampus split at y = −21; vmPFC as Brodmann areas 10, 11, 14, 24, 25, 32. Targets resampled to TR = 1.3 s and convolved with a canonical HRF. PLS decoders per region and target; representational geometry by correlating decoded category time series and agglomerative clustering; leave-one-participant-out Spearman fits of full and valence models; partial correlations controlling for decoded TEM p. Code on GitHub (eccolaboratory/EmotionConceptRepresentation), TEM from torch_tem.

Concepts

• affective space. Here: the 2D metric MDS of dissimilarity (1 − r) between the 13 categories’ consensus rating time series, with film moments projected into it by rating-weighted category coordinates (p. 10).
• map-like. Representing concepts as locations in a relational space that supports inference about transitions; operationalised as decodability of space positions (vmPFC) and of TEM units.
• transition (in the abstract). In the model, a random-walk step between adjacent nodes of the grid built from rating co-occurrence. The study does not estimate transition probabilities between emotions in the films or in anyone’s experience.

Connections

It applies TEM (Whittington et al. 2020) to emotion, after grid-like codes for abstract spaces (Constantinescu et al.; Park et al.; Nitsch et al.) and the authors’ review (ref. 43). Its affective space descends from Russell’s circumplex (ref. 1). It cites Barrett’s constructed-emotion work (refs. 3, 5, 74) and Lindquist and Satpute on the default network’s role in conceptualising emotion (refs. 72–73). It names reinforcement-learning models of hippocampus as an untested alternative (p. 8). In the record, Barrett’s theory is NOTE-434 and THEORY-057.

Bearing on the record

• How the map is learned. Not studied in people. The participants only watched; their emotion labels came from others; no change across sessions or films is analysed. The learning in the paper is the TEM’s, and the TEM learns by acting: its structural layer g “updates its representation based on the received action” by path integration (p. 9), and its agent explores by random actions. So the paper’s own model of map learning is an action-in-the-loop learner, chosen from the spatial-navigation literature, not tested against an action-free alternative (my observation). Passive exposure, judgement and action are not separated.
• Does the map shape how later episodes are read? Not tested. Decoders generalise across films, which shows a representation shared across stimuli, not one that changes how the next film is read. The paper leaves the developmental question open: “longitudinal research is needed” (p. 7).
• Constructed emotion, NOTE-434 and THEORY-057.
  • The paper puts the decisive question as open: whether the concepts “reflect the statistical regularities of important life events that are shared across individuals and cultures, or whether they are constructed more flexibly as ad hoc categories shaped by individual experience, remains an open question” (p. 7), citing Barrett (ref. 74).
  • Its closing conjecture, that the map “is the product of computations performed in hippocampal systems, rather than being an innate structure afforded by the human brain” (p. 8), is congenial to construction from experience, but it is about the valence–arousal structure, not about whether a concept constitutes the emotion.
  • Under THEORY-057’s own promote_when, this is the kind of evidence that “cannot settle it”: classifiers within one study. Leave-one-film-out generalisation is still one dataset.
  • The cross-participant hierarchy (r = .978) looks like evidence of a shared, not perceiver-specific, structure. But every participant’s decoder is fit to the same consensus ratings, so similar decoded category correlations are partly built in (my observation). It does not count against perceiver dependence.
• CLAIM-341. The registration hoped for “an observer shaped by its history, through which later emotional episodes are read”. The paper supplies a candidate substrate for such an observer, a learned relational structure separate from content, but no evidence that it is shaped by the agent’s own judgements, and no evidence that it reshapes later reading. Where it models learning, action is in the loop. So it neither supports nor undermines the claim. If anything, it is a small reminder that the standard computational account of map learning is an action–perception one.
• CLAIM-336. TEM’s factorisation, a structural code g reused across environments and a content code p bound to it by Hebbian memory, is an analogue of a context (frame) under which content is interpreted, and of a structure that persists while content changes. That is analogy, not evidence: the paper has no influence allocation and no judgement.
• CLAIM-342 and QUESTION-121. Nothing directly.
• Checks of the current LIT.
  • “hippocampal activity represented emotion concepts in a structured hierarchy and ventromedial prefrontal cortex tracked locations in a two-dimensional valence–arousal space”: correct as decodability, with small correlations, and the space is an MDS of rating co-occurrence.
  • “responses were predicted by the statistical regularities of emotion transitions across several time scales”: overstated. Brain patterns could be decoded into a TEM’s multi-scale activations; the “transitions” are random-walk steps on a grid built from rating correlations, not observed emotion transitions.
  • The decoded emotions are an independent sample’s consensus ratings, not the viewers’ own reports.
  • “Why this is here” asked whether the map is shaped by judgement, passive exposure or action: the study does not say for people; its model learns by action.
• For the anthology. No instruction for machine-learning practice.

Limitations

• Effects are small in absolute terms: prediction–outcome correlations near 0.04–0.10.
• No measure of the viewers’ own experience; ratings from another sample.
• One model (TEM); reinforcement-learning alternatives named, not fitted (p. 8).
• Entorhinal cortex did not show the predicted structure, possibly for imaging reasons (p. 7).
• Two dimensions oversimplify emotion, as the authors say (p. 8); fMRI cannot see grid or phase codes.
• Reporting oddities: p. 3 says “13 films total” for leave-one-film-out, p. 10 averages “across the 14 validation sets”; vmPFC’s d is printed equal to its Δz (0.0118), and the amygdala’s d repeats entorhinal cortex’s (−0.0195) (p. 4).
• Supplementary material was not read for this note.

Open questions

• Does the map change with exposure within the study, across the four sessions? The dataset could be asked.
• Does decoding of the viewers’ own ratings, if collected, follow the same dissociation?
• Does a map learned without action (for example by passive sequence exposure, as in successor-representation models) fit the data as well as TEM’s action-driven one? That would bear on CLAIM-341.