A Shared Value Axis Across Person, Place, and Object in Remembered Events

When people recall a personal memory, they do not value it as one undivided lump. They like the friend who was there, feel neutral about the restaurant, and dislike the rain. Brain imaging has shown that one region tracks this kind of liking. The ventromedial prefrontal cortex (vmPFC) is the region just above the eyes that is usually studied as the brain's reward calculator. Its average activity rises with how much a person likes specific items from their past (Lin and Horner 2016) and with the value of items in imagined scenarios (Lin et al. 2015). Those studies used a single number per region, its average activation. They cannot tell whether the region contains one common value code that applies equally to people, places, and objects, or separate codes for each kind of component that happen to raise average activity in the same way.

We propose to answer that question with a cross-component transfer test. We train a linear support vector machine (SVM) to separate liked from disliked people recalled from participants' own memories, using fine-grained patterns of activity across the many small volume elements (voxels) of vmPFC. A linear SVM is a classifier that finds the flat boundary separating two classes with the widest possible gap. We then test, without retraining, whether the same boundary separates liked from disliked places and objects drawn from memories the classifier never saw. Two controls make the result informative. First, we remove each pattern's average activity before classifying, so that any successful transfer cannot be explained by the single-number effect already known. Second, we check the result against a one-feature classifier that uses only average activity. If transfer survives both controls, vmPFC carries a common-currency value direction for the components of remembered events. If it does not, the known value effect is either a plain change in overall activity level or a set of component-specific codes, and the experiment says which.

Core Claim and What Is New

The specific claim is that, during autobiographical recall, the pattern of activity across vmPFC voxels encodes liking along a direction shared by social, spatial, and object components. This direction carries information beyond the region's mean activity. The linear SVM is the measuring instrument rather than the contribution. Its weight vector defines that direction explicitly, and transfer across component types is a falsifiable, out-of-sample test of "shared".

The obvious version of this intersection is to decode liked versus disliked items from vmPFC with an SVM. The research file lists this as unattempted, but it would be a weak paper on its own. A reviewer could reasonably reply that the univariate parametric effect in Lin and Horner 2016 already implies above-chance decoding from average activity alone, so a bare decoding accuracy adds little. We therefore change the question. We ask about the structure of the value code: whether it is shared across component types, whether it is multivoxel rather than an amplitude shift, and whether it is also shared between real memories and imagined scenarios. Each of these has an outcome that a univariate analysis cannot produce.

The closest decoding work is Schrouff et al. 2012. They decoded self-paced mental imagery of faces, buildings, and animals with SVMs and Gaussian processes. They did not study personal memories, subjective value, or prefrontal cortex. We reuse their main methodological lesson: self-paced internal states produce unbalanced classes that skew SVMs. The computational models of autobiographical memory are different in kind. These include AM-ART with its emotional retrieval coefficient (Wang et al. 2019), recursive kernel support vector regression as a theory of episodic storage (Leibold 2022), and cortical–medial temporal lobe consolidation circuits (Albesa-González et al. 2025). None of them measures human brain activity. They motivate the question of how value attaches to stored events but do not answer it.

Research Program

Data collection. No dataset in the file pairs brain recordings with component-level liking ratings for personal memories, so we will collect one. Before scanning, roughly 30 participants each generate about 40 autobiographical memories in an interview. For every memory we record one central person, one place, and one object, giving three components per memory. Participants rate each component for liking and also for familiarity, vividness, and emotional arousal. Lin and Horner found that familiarity explained variance in vmPFC of its own, so liking must be separated from it.

In the scanner, each trial cues a single component with a short participant-specific label, and the participant re-experiences it for a fixed interval. Separate whole-memory recall trials follow the paradigm of Lin and Horner 2016. A third block reuses the need-state imagery paradigm of Lin et al. 2015, in which items are valued within an imagined context. This block lets us test whether the value direction learned from real memories also applies to imagined scenarios. That question is pointed, because Lin and Horner reported that the autobiographical peak lies more ventral than the imagined one.

Analysis. We estimate one activity pattern per trial within an anatomically defined vmPFC region, with a searchlight analysis across the whole brain as a secondary analysis. We then subtract each pattern's mean across voxels. Labels are liked versus disliked, split at each participant's median rating within each component type.

Two features of the design need explicit handling. First, components of one event are bound together and tend to be retrieved together (Horner 2013). A classifier could therefore succeed by recognizing which memory is active rather than how much the component is liked. All cross-validation will hold out whole memories, so a test component never shares a memory with any training component. Second, liked items are likely to be over-represented. We use a cost-weighted linear SVM:

$$\min_{w,b}\ \tfrac{1}{2}\|w\|^2 + C\sum_i \frac{n}{2\,n_{y_i}}\,\xi_i \quad \text{s.t.}\quad y_i(w^\top x_i + b)\ge 1-\xi_i,\ \xi_i\ge 0.$$

Here $x_i$ is the mean-removed voxel pattern of trial $i$, $y_i\in\{-1,+1\}$ its liking label, $w$ and $b$ the boundary's direction and offset, $\xi_i$ the slack for trials that fall inside the gap or on the wrong side, $C$ the penalty on that slack, $n$ the number of training trials, and $n_{y_i}$ the number in trial $i$'s class. The weight $n/(2n_{y_i})$ gives both classes equal total influence, so the boundary is not pulled toward the majority class. This directly addresses the failure mode reported by Schrouff et al. 2012. We report balanced accuracy throughout.

The main quantity is the transfer matrix. Train on one component type and test on each of the others, then average the off-diagonal cells. In addition, we compare the cosine similarity between weight vectors trained separately on each component type against a null distribution built by shuffling labels. To check confounds, we train classifiers on familiarity, vividness, and arousal labels and measure how much liking accuracy survives when we project those directions out of the data.

Baselines

The naive reference is chance, set at 50% balanced accuracy. We estimate the null distribution empirically by permuting labels within participants and within memories, keeping the held-out-memory splits fixed. Any accuracy is judged against this null, not against the nominal 50%.

The most important baseline is a univariate classifier, and it doubles as the ablation that isolates the proposal's key ingredient. It is a logistic regression whose only input is the average vmPFC activity on each trial, which is the classification analogue of the parametric analysis in Lin and Horner 2016. It uses the same trials, the same held-out-memory splits, and the same class weighting. It is the strongest competitor because, if value is coded as a change in overall activity level, it should match the pattern classifier, and the multivoxel claim would fail. The proposal beats it only if the mean-removed SVM transfers across component types above the permutation null while this one-feature model does not, or if the SVM exceeds it by a margin that is significant at the group level in a paired test across participants. A second ablation runs the SVM on patterns with the mean left in. This shows how much of any raw accuracy comes from the average-activity component.

The strongest existing method from the file for decoding self-paced mental states is the Gaussian-process classifier, and the specific GLM plus recursive-feature-addition SVM pipeline, from Schrouff et al. 2012. On self-paced imagery they report balanced accuracy of 24.09% to 69.78% across pipelines. Their best SVM and best Gaussian-process pipelines each decoded imagery in 12 of 16 subjects. That was a three-class perceptual task, so the numbers set an expectation for how noisy self-paced decoding is rather than a target to beat. We found no reported result on Lacuna for decoding subjective value from autobiographical recall. We will run the Gaussian-process classifier on identical features and splits, with the same nested cross-validation budget for its hyperparameters as for the SVM's $C$.

A kernel SVM using a radial basis function, which allows curved boundaries, serves as the non-linear comparison, with matched tuning. Finally, a within-component SVM, trained and tested on the same component type, provides the ceiling against which transfer is read. Transfer close to that ceiling means the code is shared. Transfer far below it, with within-component accuracy still above chance, means the codes are component-specific.

Why not just…?

Why not just run the univariate parametric GLM, as the original studies did? The GLM shows that average vmPFC activity rises with liking. It cannot tell a shared multivoxel direction apart from separate codes for each component type that all raise the average. It also gives no prediction on held-out memories. This alternative is a real threat to the whole idea: if the one-feature baseline matches the mean-removed SVM, we will report that the value code in vmPFC is an amplitude signal. The comparison is built into the experiment so that this outcome is detected rather than hidden.

Why not just use a Gaussian-process classifier, since it handled unbalanced self-paced data better? In Schrouff et al. 2012 the advantage appeared under imbalance. Their best SVM and best Gaussian-process pipelines tied at 12 of 16 subjects. Our cost-weighted objective targets the imbalance directly. We also need an explicit weight vector to compute similarity between directions and to project out confounds, and the linear SVM provides one directly. We will nonetheless run the Gaussian-process classifier. If it transfers clearly better, we will adopt it and treat its implied linear direction as the measured object.

Why not just fit a non-linear or deep model, given that memory retrieval relies on non-linear pattern completion? With about 30 participants and a few hundred trials each against thousands of voxels, deep models have far more parameters than the data can constrain. The radial-basis kernel SVM is the manageable test of non-linearity. If it beats the linear model on transfer, that supports the non-linear-binding view drawn from Horner 2013 and Leibold 2022, and the claim of one shared linear axis would be weakened accordingly.

Proposed Experiments and Decisive Checks

The decisive test is the mean-removed transfer matrix in vmPFC under held-out-memory cross-validation. The core claim is supported only if average off-diagonal balanced accuracy exceeds the permutation null at the group level and the one-feature baseline fails to transfer. It is refuted if transfer sits at chance while within-component decoding succeeds, which would indicate component-specific codes. It is also refuted if the one-feature baseline transfers equally well, which would indicate an amplitude code.

A second check asks whether the direction survives after projecting out familiarity, vividness, and arousal, which addresses the open confound problem noted for Lin and Horner 2016. A third tests whether the direction learned from real memories decodes value in the imagined-scenario block, and whether it does so better in ventral or in more dorsal searchlight positions. That result bears directly on the ventral shift Lin and Horner reported.

Finally, the encoding–retrieval dissociation found for schemas by Masís-Obando and Norman 2022, using OpenNeuro ds001510, raises the possibility that value is coded outside mPFC during recall. The whole-brain searchlight transfer map tests this, in particular whether posterior medial cortex carries the shared value direction alongside vmPFC or instead of it.