Temporal and Geometric Signatures of Excitatory and Inhibitory Population Codes in Primate Object Representation

Distributed population activity in macaque inferior temporal (IT) cortex supports core object recognition, yet the contributions of excitatory (Exc) and inhibitory (Inh) neurons remain unclear. We recorded large-scale neural activity using implanted Utah arrays across IT cortex in three macaques (8 object types × 80 exemplars) and identified 124 Exc (broad-spiking) and 37 Inh (narrow-spiking) neurons via waveform clustering. We asked how the temporal evolution and geometric structure of Exc and Inh representations contribute to object coding. Inh neurons responded earlier and more reliably (Δlatency≈15 ms), achieving linearly separable stimulus representations before Exc neurons (start<100 ms: Z=3.0, p = 0.001). This early inhibitory recruitment is consistent with rapid stabilization and gain control predicted by balanced-network models, in which inhibition dynamically constrains excitatory drive before category-specific manifolds emerge. Over time (~150–250 ms), Exc and Inh populations reached comparable decoding accuracies (t(11) =0.009, p =0.993), but Exc decodes predicted monkeys’ image-level behavior significantly better (Z=67.0, p = 0.004). Geometric analyses revealed the basis of this divergence. Although Inh manifolds exhibited modest between-category radius (mean = 0.3), they showed disproportionately high within-category radius (mean = 0.78), reducing categorical separability (Z = 46.5, p=0.003). Exc manifolds were more compact both within (0.60) and across categories (0.24), yet became higher-dimensional over time (150–200 ms: Exc = 14, Inh = 11; t(24) = 9.9, p < 0.001). Thus, inhibition expands and stabilizes variance early, while excitation consolidates high-dimensional, behaviorally aligned manifolds. Consistent with this view, deep networks predict Exc responses significantly better than Inh (Z = 24.0, p < 0.001), suggesting that current models capture the excitatory backbone of IT computation while missing inhibitory control dynamics. Together, these results identify complementary temporal and geometric roles for excitation and inhibition in shaping robust visual object codes.

Recommended citation: Muzellec, S.*, Sanghavi, S.*, & Kar, K. (2026). Temporal and Geometric Signatures of Excitatory and Inhibitory Population Codes in Primate Object Representation. Cosyne.