One theory for early sensory cortex is that the brain learns the statistics of the environment, and predicts incomplete information about the environment in a Bayesian optimal manner.
The authors investigate how "optimality" of the V1 neural code evolves over time in ferrets (which are a useful model organism in vision because their eyes open after they are born, so you can tightly control their visual experience). The authors describe two ways of measuring the statistical model stored in V1. One way is to observe the spontaneous activity in the absence of stimuli. The other is to record responses of V1 neurons to natural scenes. If the model of the visual world stored in V1 really mirrors the statistics of natural scenes, then these two measurements will be the same, as measured by KL divergence.
Inexperienced ferrets had V1 spontaneous activity that diverged from the average evoked response for natural scenes, but this divergence declined over a few months to nothing. Learning of natural scene statistics was accompanied by an increase in correlated spiking activity and a decrease in sparseness. Correlations appear to be important for encoding natural scene statistics. One hypothesis in neural coding is that the ideal neural code is one where units are independent and activity is sparse, but in this experiment learning decreased both of these features.
The authors also found that V1 learns to predict future visual input based on the temporal structure of natural scenes, and that unnatural stimuli did not show the learning trend observed for natural stimuli.
- A sparse independent code stores information well, but has weak computational power.
- What does the decrease in sparseness and independence that accompanied visual learning mean ?
- Adaptation to natural statistics is harder to apply in extrastriate visual areas. These areas have significant attentional modulation, and it is difficult to separate task-dependent attention based learning from model learning within the system.