Divisive Normalization in Monkey Middle Temporal Cortex

The visual system transforms patterns of light into coherent representations of the external world through a hierarchy of specialized areas; each tuned to extract particular features of the visual scene. Maintaining response selectivity or feature tuning across this hierarchy, so that neurons remain informative despite the complexity and variability of natural inputs, is a fundamental need. Normalization, a canonical operation observed throughout the brain and across species, has long been understood as one mechanism by which selectivity is preserved. Yet normalization varies considerably in its strength, both across neurons and across stimulus conditions, and the source of this variance has remained poorly understood.

In Chapter 2, I show that this variance is systematically explained by a previously unappreciated parameter: the spatial position of stimuli within a neuron’s receptive field (RF). Stimuli that engage regions of higher RF sensitivity produce stronger normalization, a relationship captured by the RF weight at each location. This observation motivates an intensity-weighted normalization framework, in which normalization operates as a pure intensity-weighted average over all inputs, known and unknown, rather than an approximate contrast-weighted average over experimentally controlled stimuli. This reformulation has a significant conceptual consequence as it reveals that spontaneous activity contributes to normalization indistinguishable from experimentally controlled stimuli, implying that normalization is continuously active rather than recruited only by stimulus competition. This property may prove especially consequential in higher-order areas where inputs are complex and largely uncharacterized. 

In Chapter 3, I apply this framework to ask how normalization shapes the correlated variability of neuronal populations. I show that spike count correlations in macaque middle temporal area covary with normalization strength in a manner analogous to prior reports in visual area V4, suggesting that this relationship is a general property of visual cortex rather than an area-specific phenomenon. The organization of these correlations can be understood along two axes: a neuron’s direction tuning profile, which determines the sign of correlated activity, and its response magnitude, which acts as a gain that scales the correlation strength. Together, these findings reframe normalization not merely as a mechanism for controlling individual neuron responses, but as an organizing principle for the structure of population-level variability. 

Overall, in this thesis, I describe a new and simpler model of normalization that provides novel framework to evaluate how neuronal signals are integrated from their inputs.