Emergence of reliable sensory representations in cortical development
Abstract
Newborns must rapidly learn to extract information from a novel sensory environment upon birth. To guide survival-critical behaviors, their neural circuits must establish reliable internal representations of key environmental features, such as edge orientation. Yet, while cortical networks possess rich recurrent structure and organized spontaneous activity prior to eye opening, their selectivity for these features remains weak. This raises the "naïve circuit problem": how do circuits with elaborate pre-existing structure develop reliable representations upon their first encounter with structured sensory input? We address this question by viewing cortical circuits as strongly interacting systems where connectivity determines function. In such networks, external inputs aligned with the system's dominant recurrent modes are selectively amplified, while misaligned inputs fail to engage the network’s dynamics. In the mature visual cortex, orientation-tuned feedforward inputs target assemblies of neurons already preferentially interconnected by long-range horizontal connections. This feedforward-recurrent alignment is thought to enable the selective amplification necessary for strong, consistent sensory representations. We hypothesize that this alignment is not pre-specified but develops with visual experience. In ferret visual cortex, eye opening occurs thirty days after birth, allowing experimental access to circuits with established recurrent structure encountering novel visual input. We combined analysis of longitudinal mesoscale calcium imaging with computational modeling of recurrent networks to investigate how alignment determines response properties. At eye opening, individual responses evoked by oriented gratings are strong and structured, but high-dimensional and highly variable, limiting stimulus discriminability. Critically, these early evoked patterns are similar to, but distinct from spontaneous activity at the same age. Within one week of visual experience, responses become low-dimensional and highly reliable, and now correspond to concurrent spontaneous activity patterns - changes that did not occur under binocular deprivation. A minimal linear network model suggests that initial misalignment between feedforward inputs and recurrent network modes could explain these immature response properties. To test this principle in biological circuits, we developed a nonlinear network model with spatial structure. In this model, selective amplification primarily affects response reliability and stimulus-response similarity rather than overall response amplitude. The model predicts that stimulation guided by spontaneous correlation should evoke more reliable responses closer to the stimulus structure than random control stimuli. We then tested these predictions using data from patterned optogenetic stimulation in primary ferret visual cortex prior to eye opening. Both in vivo and in silico, reliability and similarity to the stimulus selectively depended on alignment with spontaneous activity. Together, these results demonstrate selective amplification of millimeter-scale activity within nonlinear cortical networks, corroborating that misalignment could underlie unreliable responses in naïve cortex. An alternative hypothesis for the emergence of reliable orientation representations is an increase in discriminability in the feedforward input alone. To disentangle these possibilities, we modeled orientation-tuned input from layer 4 to a recurrent layer 2/3. Response properties such as reliability improve with either increased input reliability or alignment. However, the tuning dynamics of sustained layer 2/3 responses depend only on the alignment. Whole-cell recordings confirmed this prediction: tuning stability during sustained responses emerged only following visual experience. Together, this suggests that both the enhanced discriminability of L4 feedforward inputs and their alignment with recurrent layer 2/3 interactions contribute to the emergence of reliable representations. To investigate what network changes could underlie the changes in alignment, we analyzed spontaneous and evoked activity in the same animals over multiple days. Spontaneous and evoked activity both reorganized toward a common manifold, with novel activity patterns emerging that were not present in naïve cortex. These results were best explained with coordinated changes in both feedforward and recurrent connections in a minimal network model. In summary, we suggest feedforward-recurrent alignment as an organizing principle for reliable cortical representations. This offers a framework for understanding both typical development and the consequences of abnormal early experience. The framework demonstrates how self-organizing systems can rapidly acquire function when encountering novel structured input, with potential implications for artificial learning systems and developmental interventions.