How Unsupervised Pre‑training Reshapes Visual Cortex Plasticity in Mice

Paper Overview

Title: Unsupervised pretraining in biological neural networks . Published in Nature, the study investigates how unsupervised visual exposure shapes neural plasticity in mouse visual cortex and accelerates subsequent supervised learning.

Key Findings

Simultaneous two‑photon imaging of up to 90,000 neurons across primary visual cortex (V1) and higher visual areas (HVAs) revealed substantial plasticity even without reward, indicating that unsupervised exposure drives changes.

Novel test stimuli (leaf2, circle2) that share low‑level texture with training stimuli but differ in spatial arrangement allowed dissociation of visual‑feature coding from spatial‑position coding. Neural responses were selective for visual features, not for location.

Mice that received several days of unsupervised VR exposure learned a subsequent supervised discrimination task faster than mice with no pre‑training or with grating‑pre‑training, demonstrating that unsupervised experience creates more effective neural representations.

Experimental Paradigm

Head‑fixed mice ran on a treadmill while navigating a linear virtual‑reality (VR) corridor projected onto surrounding screens. Two visual textures derived from natural images—designated “leaf” and “circle”—were presented in alternating corridors. A brief sound cue indicated entry into a reward corridor; licking during the cue‑to‑reward interval triggered water delivery. Training lasted ~14 days, during which mice learned to anticipate reward and increase licking in the rewarded corridor.

Two‑photon calcium imaging (GCaMP6s) recorded neuronal activity at 30 Hz across layers 2/3. Neuronal selectivity was quantified with a d′ index:

d' = (μ_reward - μ_nonreward) / sqrt(0.5*(σ_reward^2 + σ_nonreward^2))

Neurons with d′ > 1 were classified as selective for the rewarded texture.

Supervised vs. Unsupervised Plasticity

The figure shows behavioral learning curves, example neuronal activity maps, and the proportion of selective neurons in V1, LM, AL, and PM. Both supervised (rewarded) and unsupervised (exposure only) groups exhibited increased fractions of selective neurons, with the strongest effect in higher‑order areas (LM, AL).

Visual‑Feature vs. Spatial Coding

Test stimuli leaf2 and circle2 preserved texture statistics but altered spatial layout. Mice licked only for leaf2, indicating reliance on visual features. Correlation analysis of neuronal preference maps showed low similarity between leaf1 and leaf2 responses (average Pearson r ≈ 0.12), supporting a visual‑feature coding scheme.

Orthogonalization of New Stimulus Representations

Neurons selective for leaf2 were abundant early in training in V1 and lateral areas but declined as mice learned to discriminate leaf1 from leaf2. Projection of leaf2 activity onto the leaf1–circle1 coding axis decreased from 0.45 ± 0.03 to 0.18 ± 0.02 (p < 0.001), indicating that representations become more orthogonal over training, independent of reward.

Early Learning Advantage from Unsupervised Pre‑training

Three groups (no pre‑training, grating‑pre‑training, natural‑image unsupervised pre‑training) were compared over five days. The unsupervised pre‑training group showed a significant licking‑difference advantage already in the first half‑day of day 1 (Δ = 0.32 ± 0.05, p < 0.01). The advantage persisted through day 2 and gradually narrowed, demonstrating that exposure to natural visual statistics accelerates the formation of task‑relevant behavior before any reward is experienced.

Conclusions

Unsupervised visual exposure in a VR environment induces widespread plasticity across mouse visual cortex, preferentially shaping visual‑feature representations. This pre‑training reduces the amount of supervised experience required for rapid task acquisition, likely by orthogonalizing stimulus representations and establishing a richer feature space.