Visual cortex neurons do more than fire harder: in macaque V1 and V4, a second feature pushes activity below baseline
Synopsis
Many neurons in macaque V1 and V4 do not merely raise firing for a preferred stimulus; they are excited by one visual feature and suppressed by a distinct one around an elevated baseline. Using functional digital twins, deep networks trained on recorded responses that predict how a neuron will answer new images, researchers identified each neuron's most and least activating images across more than a million natural images, and confirmed that these images sit at the extremes of recorded responses and hold up under independent models. The same bidirectional selectivity appears in mouse V1 and lateral visual areas.
Interpretation
A substantial share of visual cortical neurons respond in two directions: around an elevated baseline firing rate, one visual feature enhances activity and a different feature suppresses it. Neurons were traditionally treated as feature detectors that raise firing for a preferred stimulus and stay largely unresponsive otherwise, with inhibition viewed as broadly tuned gain control unrelated to specific features. Recordings in macaque V1 and V4 show response sparsity along a continuum, with higher baseline firing associated with less sparse responses; for non-sparse neurons, the model-predicted most activating image matched a high recorded response and the predicted least activating image matched a response below baseline.
This bidirectional selectivity collapses onto a low-dimensional axis: activity rises as an image approaches the most activating image in perceptual similarity space and falls as it approaches the least activating image. How excitation and suppression jointly shape tuning over natural images was poorly characterized, because recording time limits how many stimuli can be shown. Anchoring a two-dimensional space on each neuron's most and least activating images and embedding about 200,000 natural images with DreamSim, a perceptual similarity embedding aligned with human judgments, non-sparse V4 responses varied smoothly along the diagonal; replacing the anchor images with random images significantly reduced explained variance.
Most and least activating images are shared across the population: an image that strongly drives one neuron is also likely to strongly excite or suppress other neurons in the same area. If each neuron's preferred and suppressive features were independent, one neuron's extreme images should land randomly in the response distributions of other neurons. In V1 and V4, one neuron's most activating images produced right-skewed response percentiles in other neurons and its least activating images produced bimodal distributions, while random images gave near-uniform distributions; extreme images from one macaque elicited similar patterns in neurons from another.
Perspective
The conclusions cover spiking recordings from macaque V1 and V4 (V1 data from a previously published study, V4 newly collected) and from mouse V1, LM and LI, and the analyses focus mainly on neurons the models classify as non-sparse; least-activating-image analyses are restricted to these neurons because sparse neurons have low baselines and little measurable suppression.
The authors note that conclusions rely heavily on in silico analysis rather than direct experimental validation, especially on the suppressive side; the analyses do not determine which interneuron subtypes provide the inhibitory input; the perceptual similarity space DreamSim is based on human judgments while the neural data come from macaques; and the feature relationship between most and least activating images is not yet quantified, with no systematic opponency in dimensions such as orientation observed.