There is a discrepancy between the generally accepted role of ongoing activity during visual development, where spontaneous firing is viewed as an important guiding activity indispensable for proper emergence of the visual structure, and during visual perception, where spontaneous neural activity is considered to be unwanted noise. This discrepancy stems from the presently dominant view which posits that visual information is analyzed in a feedforward signal-processing manner where ongoing activity is accidental and can be neglected. To study this discrepancy, we analyzed multi-electrode recordings in the primary visual cortex of awake behaving ferrets (N=20) at postnatal day (P) 24-26, P44-45, P71-90 and P131-168. Multi-unit recordings were obtained in three different conditions: in the dark, when the animals watched random noise sequences, and when they saw a natural movie. At all ages there was a significant spatio-temporal structure in the observed neural activity and this structure showed a distinctively evolving pattern across ages. The high spatial correlations across different recording sites during the dark condition ruled out the possibility of averaging out the noise correlations and thus questioning the validity of feed-forward signal processing models. An alternative model is based on a generative Bayesian framework where ongoing activity represents momentary perceptual biases of the brain based on previously obtained information and internal states. To test the validity of this framework, the same data was analyzed using a Hidden Markov Model. We found clearly distinct internal states in all conditions defined by approximately stationary firing rates and abrupt transitions between states. The identified HMMs were specific to particular conditions classifying untrained neural activity correctly about 90% of the times. These findings suggest that even in the primary visual cortex neural processing can be best described as a rapid dynamic transition between a large number of states, where the external input modulates the intrinsic dynamics by selectively boosting particular states.