VIP Inhibitory Neurons in the Visual Cortex Perform Two Types of Predictive Processing: Stimulus Specific & Non-specific

VIP inhibitory neurons in the cortex receive neuromodulatory and top-down feedback inputs, hence they are well-suited to process behavioral and predictive signals. A recent study found that VIP neurons respond to unexpected image omissions in an image sequence (Garrett et al., 2020). This omission-evoked activity may represent changes in the animal’s behavioral state (e.g., running or arousal), or it may represent a prediction-error signal that could either be stimulus specific, as suggested by the classical predictive coding models (Rao & Ballard, 1992), or stimulus non-specific. We investigated this question by training two Generalized Linear Models (GLMs) on spiking activity extracted from calcium imaging data from the visual cortex of behaving mice (Allen Institute Visual Behavior 2p Dataset). Both GLMs included the same image, behavioral and task kernels, but differed on having either distinct image-specific omission kernels, or a single image-non-specific omission kernel. First, both GLMs demonstrated that the behavioral features do not contribute to the VIP response around omissions; instead, VIP response was largely driven by omission features, representing a potential prediction-error signal. Second, the non-specific-omission GLM provided a better fit to the activity of most, but not all, VIP cells, suggesting that some VIP cells may carry an image-specific prediction signal. We directly tested this hypothesis by computing the image selectivity of omission responses. Indeed, some VIP cells carry image-specific signals about omissions. Overall, our findings suggest that VIP activity following unexpected image omissions is not due to behavioral changes; instead, it may indicate predictive processing in VIP neurons, which is non-specific in most cells, but stimulus-specific in a smaller subset of cells. Further studies will reveal how these distinct functional subtypes of VIP neurons differ in their inputs, outputs, and transcriptomics profile, also how distinctly they may contribute to the processing of internal, sensory, and behavioral signals.