Neural Noise- An inception-loop framework maps neuronal invariances in mouse V1, revealing bipartite receptive-field organization linked to segmentation and a synaptic-level hierarchy of increasing invariance supported by the MICrONS dataset (Nature Neuroscience, 2026).1
- Relevant for computational neuroscience and neural data analysis.1
- The study by Paul G. Fahey et al. (published Feb 25, 2026, Nat Neurosci 29:851–863) used two-photon calcium imaging of V1 L2/3 excitatory neurons in awake, head-fixed mice; a CNN model trained on responses to 5,100 natural images (median noise-normalized correlation = 0.71 across 33,714 neurons) was used to synthesize varied exciting inputs (VEIs) — dissimilar images constrained to elicit ≥85% of each neuron’s maximal response.1
- VEIs revealed a novel bipartite invariance: one receptive-field subfield encodes a fixed low-frequency spatial pattern, while the other responds robustly to random crops of a high-frequency texture, parameterized by a bipartite invariance index (BII); simulated simple cells, complex cells, and V1 neurons exhibited BII medians of 0.53, 0.87, and 0.65, respectively.1
- The variable (texture) subfield boundaries align with object boundaries defined by spatial-frequency discontinuities — screened across >41 million natural image crops and >1 million CUB bird dataset crops; 99.1% of V1 neurons preferred images containing spatial-frequency-defined object boundaries over single grating images.1
- In vivo validation confirmed VEIs drove target neurons at 74–75% of MEI activation (close to the model’s predicted 85%); V1 population activity decoded pairs of VEIs with median 80% accuracy (chance 50%), showing single-neuron invariances correspond to perceptually accessible image transformations.1
- Using the MICrONS functional connectomics dataset (>75,000 neurons, dense EM reconstruction of synaptic connectivity in ~1 mm³ of mouse visual cortex), the study found a functional invariance hierarchy in V1 L2/3: postsynaptic neurons exhibited greater diversity/invariance than their presynaptic inputs, while neurons with lower invariance formed more synaptic connections (synapse conversion rate decreased exponentially with increasing presynaptic diversity index).1
- The “like-to-like” connectivity principle — neurons with similar response properties preferentially synapse — was confirmed at synaptic resolution (not merely a byproduct of spatial proximity), with connected pairs showing higher MEI and VEI representational similarity than anatomically proximate but unconnected controls.1
- The inception-loop methodology is robust across model architectures, synthesis conditions, and recording modalities (two-photon calcium imaging and Neuropixels electrophysiology), offering a scalable tool for mapping high-order neuronal invariances in any brain area.1
- Broader implication for visual neuroscience: mouse V1 neurons appear tuned to detect texture-defined object boundaries, suggesting an early cortical contribution to segmentation that precedes higher visual areas — challenging the classic view of V1 as purely a low-level edge detector.1