Shiming Tang

Population-level measurements portray object representations in primate inferotemporal cortex (IT) as smooth, low-dimensional, and predictable by deep neural networks (DNNs). However, it remains unclear whether this structured population-level picture is representative of the full diversity of its constituent neurons. Here, we used Neuropixels 2.0 probes to record large populations of well-isolated units from an fMRI-localized face patch in macaque anterior IT while monkeys viewed more than 3,000 natural images. Single-unit responses were substantially sparser and more heterogeneous than multi-unit activity (MUA). Sparse neurons collectively formed a higher-dimensional code, supported efficient image identification, and responded later than broadly tuned neurons. Partly independently of sparseness, some neurons exhibited reliable feature randomness: their stimulus preferences were reproducible across repeated presentations but discontinuous across DNN feature spaces, rather than reflecting trial-to-trial response variability or noise. These neurons were virtually uncorrelated with the surrounding population despite being located within the same face patch. By contrast, MUA responses were denser, more correlated, lower-dimensional, and better predicted by DNNs. Together, these findings suggest a dual coding architecture in anterior IT: shared low-dimensional structure supports category generalization, whereas sparse and reliably feature-random single-neuron responses expand the representational space and support efficient identification of individual visual inputs. Response sparseness and reliable feature randomness may therefore constitute fundamental computational resources for object recognition.

Continuous flash suppression (CFS), in which a dynamic masker presented to one eye suppresses awareness of a stimulus in the other eye, is widely used to study visual subconsciousness. Although some studies report preserved high-level processing under CFS, these effects have been increasingly questioned and may partly reflect residual low-level feature processing. A key unresolved issue is how strongly neuronal responses in V1, where inputs from the two eyes first converge, are affected by CFS, and how much the remaining signals can support downstream processing. Here, we used two-photon calcium imaging to record large populations of V1 neurons in awake, fixating macaques while presenting grating stimuli under CFS. CFS strongly suppressed V1 orientation responses in an ocular-dominance-dependent manner, nearly abolishing responses in neurons preferring the masker eye or both eyes, and significantly reducing responses in neurons preferring the grating eye. Modeling analyses further indicated that V1 population activity under CFS may still support coarse orientation classification but not accurate stimulus reconstruction. These results suggest that CFS substantially degrades orientation information in V1. The residual signals may support limited low-level processing but are likely insufficient for downstream higher-level visual and cognitive tasks.

Neural networks underlie complex brain information processing, yet the role of their single-neuron topology in governing computation and behavior remains unclear, particularly regarding how it shapes individual neuron function and activity evolution during learning. Using two-photon calcium imaging, we tracked functional connectivity in thousands of posterior parietal cortex neurons as monkeys learned sensorimotor associations across days. We identified small-world networks with densely connected hub neurons that dominated encoding key task variables, driving local dynamics and neural encoding evolution during the monkeys’ task performance and learning. Dynamic transitions in hub/non-hub status captured how inter-neuronal interactions shaped neuronal encoding evolution during association formation. Modular structures supported specialized neuron ensembles, enabling segregated representations and interactions within local networks. Importantly, small-world network properties predicted behavioral performance, with global information processing efficiency increasing as learning progressed. These findings reveal how single-neuron-resolution brain networks, through small-world organization, orchestrate both global and modular neural computations within local network to mediate behavior and shape learning.

Continuous flash suppression (CFS), a dynamic masker presented to one eye suppressing the conscious perception of a stimulus shown to the other eye, has been extensively used to study visual consciousness. Various studies reported high-level visual and cognitive functions under CFS, which, however, has more recently been questioned and at least partially attributed to low-level stimulus properties. A key but unsettled issue is the extent to which the responses of V1 neurons, where inputs from two eyes first merge, are affected, as severely suppressed V1 responses would not sustain high-level processing. Here, we used two-photon calcium imaging to record the responses of large samples of V1 neurons to a grating stimulus under CFS in awake, fixating macaques. The results revealed that CFS substantially suppressed V1 orientation responses. Ocularity-wise, it nearly completely eliminated the orientation responses of V1 neurons preferring the masker eye or both eyes, while also significantly suppressing the responses of those preferring the grating eye. Modeling analyses suggest that, under CFS, the brain retains the ability of classifying coarse orientations, but may become less capable of reconstructing the grating stimulus. Consequently, while CFS-suppressed orientation information still supports low-level orientation discrimination, it may not suffice for high-level visual and cognitive processing.

Monitoring behavioral outcomes is essential for optimizing behavior and learning new tasks. However, how the brain monitors outcome information during long-term learning remains largely unknown. Using two-photon calcium imaging in behaving monkeys, we tracked neuronal activity in area 7a of the posterior parietal cortex (PPC), an important associative cortex that has not been typically implicated in associative learning (AL), while the macaque monkeys learned new visuomotor associations across multiple days. We found robust neuronal representation of salient behavioral outcomes (correct vs. incorrect) which closely correlated with the monkeys’ learning behavior. Furthermore, these outcome representations significantly reorganized when monkeys transitioned to learn novel associations and gradually evolved over subsequent learning days, a process constrained by the functional connectivity among neurons within local network. These results suggest a substantive role for primate PPC in long-term AL through monitoring trial outcome, and indicate a principle for long-term learning: network connectivity constrains the evolution of neural encodings.

Our brain is remarkably limited in how many items it can hold simultaneously, but it can also represent unbounded novel items through generalization. How the brain rationally uses limited resources in working memory (WM) remains unexplored. We investigated mechanisms of WM resource allocation using calcium imaging and electrophysiological recording in the prefrontal cortex of monkeys performing sequence WM (SWM) tasks. We found that changes in the neural representation of SWM, including geometry, generalizable and separate rank subspaces, reflected WM load. SWM resources, represented by neurons’ signal strength and spatial tuning projected onto each rank subspace, were shared flexibly between ranks. Crucially, the prefrontal cortex dynamically utilized shared tuning neurons to ensure generalization, while engaging disjoint and spatially shifted neurons to minimize interference, thus achieving a trade-off between behavioral and neural costs within capacity. The allocated resources can predict monkeys’ behavior. Thus, the geometry of compositionality underlies the flexible use of limited resources in SWM.

Neural networks underlie complex brain information processing, yet the role of their single-neuron topology in governing computation and behavior remains unclear, particularly regarding how it shapes individual neuron function and activity evolution during learning. Using two-photon calcium imaging, we tracked functional connectivity in thousands of posterior parietal cortex neurons as monkeys learned sensorimotor associations across days. We identified small-world networks with densely connected hub neurons that dominated encoding key task variables, driving local dynamics and neural encoding evolution during the monkeys’ task performance and learning. Dynamic transitions in hub/non-hub status captured how inter-neuronal interactions shaped neuronal encoding evolution during association formation. Modular structures supported specialized neuron ensembles, enabling segregated representations and interactions within local networks. Importantly, small-world network properties predicted behavioral performance, with global information processing efficiency increasing as learning progressed. These findings reveal how single-neuron-resolution brain networks, through small-world organization, orchestrates both global and modular neural computations to mediate behavior and shape learning.

Continuous flash suppression (CFS), a dynamic masker presented to one eye suppressing the conscious perception of a stimulus shown to the other eye, has been extensively used to study visual consciousness. Various studies reported high-level visual and cognitive functions under CFS, which, however, has more recently been questioned and at least partially attributed to low-level stimulus properties. A key but unsettled issue is the extent to which the responses of V1 neurons, where inputs from two eyes first merge, are affected, as severely suppressed V1 responses would not sustain high-level processing. Here, we used two-photon calcium imaging to record the responses of large samples of V1 neurons to a grating stimulus under CFS in awake, fixating macaques. The results revealed that CFS substantially suppressed V1 orientation responses. Ocularity-wise, it nearly completely eliminated the orientation responses of V1 neurons preferring the masker eye or both eyes, while also significantly suppressing the responses of those preferring the grating eye. Modeling analyses suggest that, under CFS, the brain retains the ability of classifying coarse orientations, but may become less capable of reconstructing the grating stimulus. Consequently, while CFS-suppressed orientation information still supports low-level orientation discrimination, it may not suffice for high-level visual and cognitive processing.