Neural representational geometries reflect behavioral differences in monkeys and recurrent neural networks

Despite standardized training and environments, animals can adopt different strategies to solve the same task, sometimes yielding indistinguishable performance. Understanding these latent strategies requires analyses that bridge neural population representations and behavior at the individual level. Representational geometry—the arrangement and distances among condition-evoked population activity states—can reveal structured formats that support generalization (abstract or disentangled representations) and may relate to behavior. Here, we test whether subject-specific differences in neural representational geometry in dorsolateral prefrontal cortex (PFdl) relate to behavioral differences during a visually cued stay/shift rule task. We hypothesize that differences in abstract encoding of task variables (e.g., cue shape vs. rule) will correspond to systematic differences in reaction time patterns, despite similar overall accuracy.

Prior work shows that neural population geometry can support linear separation and generalization, with abstract (disentangled) representations enabling cross-condition generalization. Such abstractions have been reported across multiple brain areas, including prefrontal cortex and hippocampus. However, direct links between representational geometry and behavior at the individual level remain limited. The need for fine-grained behavioral analyses to interpret neural data has been emphasized, and recent machine learning studies highlight how different training regimes and curricula can yield divergent internal representations while achieving similar performance. This context motivates examining geometry-behavior correspondences in single subjects and leveraging RNNs to probe mechanistic origins.

Subjects and recordings: Two male rhesus monkeys (Macaca mulatta, 10–11 kg) were trained on a visually cued rule-based task. Single-unit activity was recorded in dorsolateral prefrontal cortex (PFdl) using up to 16 electrodes; recording sites were localized via MRI and histology. Inclusion criteria required at least 5 complete and correct trials per condition. Final datasets: Monkey 1, 205/289 neurons (71%); Monkey 2, 188/262 neurons (72%). Task: On each trial, monkeys fixated (1.5 s), viewed a visual cue (0.5 s), maintained fixation through a variable delay (1.0–1.5 s), then, upon go, made a saccade to one of two targets. Four cues instructed two rules: stay (vertical gray rectangle or yellow square) or shift (horizontal gray rectangle or purple square). Correct responses were determined by combining the previous response with the cued rule. Neural data construction: Pseudo-simultaneous population trials were built by sampling spike counts (200 ms bins, 20 ms step) across neurons from trials within the same condition, generating 100 pseudo-trials per condition per time bin. Variables and dichotomies: Eight task conditions (combinations of previous response, rule, current response, and cue shape) define 35 balanced dichotomies, including four interpretable task variables (previous response, rule, current response, shape). Decoding: Linear SVM (regularization 10^3) trained/tested on pseudo-trials with 100 iterations and cross-validations. Statistical significance assessed via 100 label-shuffle null models; chance interval defined as ±2 SD around 50%. Abstract format (CCGP): Cross-Condition Generalization Performance measured by training on three conditions per side of a dichotomy and testing on the held-out conditions (16 train/test splits per dichotomy; 10 CVs each). A custom null model preserved decodability but disrupted geometry via discrete rotations/permutations of noise clouds; significance via ±2 SD around 50%. Shattering dimensionality (SD): Estimated as average linear decoding performance across all balanced dichotomies, reflecting the capacity to linearly separate many input-output mappings. Geometry visualization: Metric multi-dimensional scaling (MDS) applied to a dissimilarity matrix of Euclidean distances between condition means, normalized by variance along inter-condition axes; also applied to single pseudo-trial data (for pairwise decoding with denoising). Behavioral analyses: Combined sessions (Monkey 1: 65; Monkey 2: 77). Reaction time (RT) defined as time from go cue to target acquisition; outliers >3 SD removed. Performance and RTs compared across rule (stay/shift) and shape (rectangle/square) via chi-square (performance) and Mann-Whitney U tests (RT). A multi-linear regression predicted single-trial RT using binary factors (rule, previous response, shape) and their interactions; 100 subsampled models fit via OLS; weights compared across monkeys via Mann-Whitney U. RNN modeling: 80 vanilla RNNs trained with proximal policy optimization (PPO). Inputs: fixation (scalar), shape (3 one-hot), color (3 one-hot), previous response (2 one-hot). Inputs projected via fixed random weights to a 100-unit ReLU expansion layer; then to 100 recurrent units (τ=100 ms; noise added). Outputs: 3-unit policy (fixate/left/right) via softmax and scalar value function. Training stopped when networks achieved ≥99% complete trials within 1500 ms of decision and ≥90% correct on a 10,000-trial validation batch. Post-training, recurrent activity analyzed on 10,000 test trials. Geometry metrics: Δ-decoding and Δ-CCGP defined as (shape − rule) during cue. Behavioral metric: ΔRT = |RT(shape difference)| − |RT(rule difference)|, where positive indicates more shape-dependent RTs. Correlations assessed between geometry metrics, training duration (trials to criterion), and ΔRT.

- Similar task accuracy but different RTs: No significant difference in overall performance across monkeys (chi-square p=0.93). Average RTs differed significantly (Mann-Whitney U p=10^-15). - Distinct representational geometries in PFdl: • Monkey 1: During cue presentation, most dichotomies decodable; shape shows highest CCGP (abstract), followed later by current response. Rule and previous response decodable but not abstract during cue; rule becomes abstract after cue offset. Higher shattering dimensionality than Monkey 2. • Monkey 2: During cue, rule and current response decodable; shape and previous response not decodable. Rule exhibits highest CCGP (abstract) during cue; current response also abstract; shape and previous response not abstract. • Both monkeys: Current response becomes abstract during cue presentation; previous response not abstract during cue. - Control for recording sites: Monkey 1 dorsal-only analyses (matching Monkey 2 sites) yielded comparable decoding results; previous response signals stronger ventrally in Monkey 1, suggesting potential ventral contributions in Monkey 2 were unrecorded. - Shape decodability with denoising: After MDS-based denoising and pairwise classification, shape is decodable under both rules in both monkeys; accuracy lower in Monkey 2, indicating both encode shape but with different geometries. - MDS visualization: Monkey 1 clusters conditions by shape and current response in an abstract format; Monkey 2 clusters by rule and current response abstractly; previous response not abstract in either. - RT patterns mirror geometry: • Monkey 1: RT differs by shape (p=0.002) regardless of rule (p=0.05). • Monkey 2: RT differs by rule (p=10^-10) regardless of shape (p=0.28). • Multi-linear RT model: Rule weight larger in Monkey 2 than Monkey 1 (p=10^-34); shape weight larger in Monkey 1 (p=10^-4). Interaction of previous response × rule strongest in both. - RNNs reproduce geometry-behavior links: • Across 80 RNNs, Δ-decoding and Δ-CCGP (shape − rule) negatively correlate with training duration (Pearson ρ≈-0.52, p≈10^-6; and ρ≈-0.48, p≈10^-5), indicating faster-trained networks emphasize shape over rule. • Geometry-behavior correlation: Δ-decoding and Δ-CCGP positively correlate with ΔRT (ρ=0.35, p=0.001; ρ=0.39, p=10^-4), linking stronger shape geometry to more shape-dependent RTs. • Example networks: NET 1 (strong shape, abstract shape/current response) shows RT differences by shape; NET 2 (strong rule, abstract rule/current/previous responses) shows RT differences by rule. Both achieve >90% accuracy. - Additional modeling insights: Kinematic analyses of population trajectories show that stronger decoding for a variable in activity space corresponds to larger separations in velocity-trajectory space for that variable, consistent with RT differences.

The study demonstrates that individuals can solve the same task with similar accuracy while employing different strategies, reflected in distinct neural representational geometries in PFdl. Monkey 1 exhibits a more "visual" strategy (abstract encoding of cue shape, higher dimensionality), compatible with a lookup-table-like mapping, whereas Monkey 2 shows a more "cognitive" strategy (abstract rule representation), potentially favoring generalization when cue sets change. These neural formats predict fine-grained behavioral differences: reaction times cluster by the variable encoded abstractly (shape for Monkey 1, rule for Monkey 2), even when overall accuracy is matched. RNNs trained with reinforcement learning recapitulate these patterns and suggest a mechanistic link between training duration and representational format: faster training preserves and leverages input-driven disentangled shape signals (lazy regime), while longer training emphasizes task-relevant abstractions like rule (rich regime). This supports a mapping from representational geometry to behavior and implies that training history or curriculum can drive strategy differences in both artificial and biological systems. The approach provides a systematic, interpretable framework to relate multi-dimensional neural geometry to individual behavioral phenotypes.

This work introduces a neural-geometry-based framework to detect and interpret individual strategy differences in a rule-based task. Despite similar performance, two monkeys showed distinct PFdl representational geometries—one abstractly encoding shape, the other rule—predicting corresponding reaction-time patterns. RNN models reproduced these geometry-behavior relationships and implicated training duration as a driver of strategy formation. These findings establish representational geometry as a practical tool to link neural coding formats to behavior and to infer latent strategies. Future work should include longitudinal neural recordings during learning, explicit behavioral generalization tests (e.g., new cue sets), and refined models that better match memory demands to test how abstract representations facilitate generalization and how training curricula shape neural geometry and behavior.

- Sample size limited to two monkeys restricts generalizability. - No neural data collected during training; strategy evolution over learning is inferred indirectly and via models. - Recording coverage differences (ventral PFdl not recorded in Monkey 2) may underrepresent some variables (e.g., previous response) despite dorsal-site controls in Monkey 1. - Reaction-time differences, though significant, are modest in magnitude. - RNN simplifications: previous response provided explicitly at cue onset (no working-memory requirement), and absolute reaction times not fitted to match animals; model architectures and input encodings may bias representational formats. - CCGP and shattering dimensionality are indirect measures that depend on decoding choices and noise modeling.