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

Laboratory tasks, despite their simplicity and controlled environments, can elicit diverse strategies in animals. Even with standardized protocols, consistent behavioral replication across laboratories remains challenging, suggesting that underlying strategies might differ. This study focuses on a visually cued rule-based task where monkeys choose between spatial targets based on a rule (stay or shift) cued by a visual stimulus. Traditional analyses often mask individual differences by averaging across subjects with varying strategies. Therefore, the research question is whether differences in neural representational geometry can reveal the underlying strategies used by individual subjects. The study is important because it addresses the challenge of inferring individual strategies from neural data, leading to a more precise understanding of brain-behavior relationships. This would improve our ability to interpret behavioral differences, and potentially to predict responses to novel tasks and conditions. The current study addresses this challenge by focusing on the geometry of the neural representations to identify the underlying strategies.

Previous research has highlighted the diversity of neural responses in cognitive areas, often appearing disorganized at the single-neuron level. However, population-level analysis reveals informative structures. Recent work has shown that representational geometry, defined by distances between points representing different experimental conditions in neural activity space, holds significant computational implications. High-dimensional representations allow for flexibility in downstream processing, while low-dimensional, abstract representations exhibit generalization properties. Abstract representations have been observed in various brain regions, but their behavioral links remain unclear. Krakauer et al. (2017) emphasize the need for a framework mapping neural data to behavior with similar levels of detail, advocating for fine-grained behavioral analysis to understand brain-behavior relationships.

Two male rhesus monkeys performed a visually cued rule-based task involving choosing between two targets based on a stay or shift rule indicated by visual cues (rectangles or squares of different colors). Single-unit recordings were obtained from their PFdl. The monkeys performed the task with high accuracy. The representational geometry was analyzed using three key aspects: (1) Linear decoding accuracies for task-relevant variables (previous response, rule, current response, cue shape); (2) Cross-condition generalization performance (CCGP), measuring the ability of linear decoders trained on a subset of conditions to generalize to unseen conditions, indicating the abstractness of representations; and (3) Shattering dimensionality, assessing the overall dimensionality of the representation. These aspects were examined during cue presentation and subsequent periods. For behavioral analysis, reaction times (RTs) were analyzed for each condition and a multi-linear regression model was fitted to predict RTs based on rule, previous response, and cue shape, including interaction terms. To provide mechanistic insights, multiple RNNs were trained to perform the same task using deep reinforcement learning. The RNNs’ representational geometries, reaction times, and training duration were analyzed to find correlations with the monkey data. Multi-dimensional scaling (MDS) was used to visualize the representational geometries in low-dimensional space.

Despite similar overall task performance, the two monkeys exhibited dramatically different representational geometries in their PFdl. Monkey 1 showed a more "visual" representation, with cue shape being strongly represented in an abstract format during cue presentation. Monkey 2 showed a more "cognitive" representation, with the rule strongly represented abstractly. These differences correlated with significant differences in RT patterns: Monkey 1's RTs varied with cue shape, while Monkey 2's varied with the rule. RNN models trained on the same task replicated these findings. RNNs reaching high performance quickly had geometries similar to Monkey 1 (strong shape representation, shorter RTs influenced by shape), while those requiring longer training resembled Monkey 2 (strong rule representation, shorter RTs influenced by rule). The amount of training needed to reach a performance criterion was negatively correlated with the difference in decoding accuracy and CCGP between shape and rule. The difference in decoding accuracy and CCGP correlated positively with the difference in reaction times (ART). Analysis of kinematic properties of the recurrent population activity in the RNNs revealed a significant correlation between representational geometry in the activity space and kinematic measures such as distance between centroids of conditions based on shape and rule in the kinematic space.

This study demonstrates that analyzing neural representational geometry reveals subtle behavioral differences not apparent through traditional performance metrics. The observed differences in representational geometries (visual vs. cognitive) in the monkeys are linked to distinct RT patterns, suggesting different strategies. The RNN models further support these findings by showing that the training duration affects the development of specific representational geometries and the resulting reaction time patterns. The more efficient networks (faster learning) developed more visual strategies and had reaction times determined by the shape. Less efficient networks (longer training) developed more cognitive strategies and had reaction times determined by the rule. The study suggests that abstract representations, as revealed by CCGP, are linked to task performance and potentially to generalization. Future studies should focus on the relationship between the amount of training and the development of specific strategies.

This study demonstrates that analyzing neural representational geometry offers a powerful tool for understanding individual differences in behavioral strategies. The striking correlation between representational geometry, training duration, and reaction times in both monkeys and RNNs suggests a mechanistic link between learning processes and strategy selection. Future research should explore the role of abstract representations in generalization to novel stimuli and investigate the effects of training paradigms on the development of different strategies.

The study is limited to two monkeys. While the RNN models provide mechanistic insights, they are simplified representations of the brain, neglecting various complexities such as the previous response retrieval from memory. The models did not perfectly replicate the monkey RT values; the focus was on the pattern of RT differences.