INTRINSIC DIMENSIONALITY OF NEURAL SIGNALS SHAPES COMPUTATION IN DOWNSTREAM CIRCUITS

To generate movement, the brain transforms sensory and cognitive information into motor commands through interconnected regions. Neural signals between regions are carried by projection neuron populations whose number defines the signal's embedding dimensionality. Temporal correlations reduce its effective degrees of freedom to a lower intrinsic dimensionality. Large-scale recordings show that in motor cortex, these correlations are so pronounced that dynamics unfold along a low-dimensional manifold. Yet whether this low dimensionality serves a functional purpose for downstream circuits, including spinal cord, brainstem, and cerebellum, remains unknown. Here, we used computational modeling to investigate how signal dimensionality shapes processing in recurrent spiking networks (Fig. 1a-i). We show that network parameters shape spontaneous dynamics (Fig. 1a-ii) and modulate encoding of one-dimensional inputs, quantified using linear readouts on evoked activity (Fig. 1b). Next, we constructed signals with fixed embedding but varying intrinsic dimensionality (Fig. 1c). Using linear readouts, we quantified encoding fidelity as the accuracy of reconstructing the inputs, and transformation capacity as the accuracy of producing a distinct target output. Results reveal that low intrinsic signal dimensionality enhances encoding fidelity (Fig. 1d-i), robust across network parameter regimes (Fig. 1d-ii). Preliminary results suggest high intrinsic signal dimensionality increases transformation capacity (Fig. 1e), pointing to a dimensionality-dependent trade-off. These findings provide a rationale for the low dimensionality of motor cortical output to downstream regions for stable motor commands. Together, they predict that signals from frontal and sensory cortices to motor cortex exhibit high intrinsic dimensionality, consistent with the demands of generating diverse spatiotemporal motor patterns.