Abstract
Top-down feedback from prefrontal cortex (PFC) can enhance the gain of feature selective neurons in early sensory areas that are tuned to behaviorally relevant stimuli (termed feature-based attention). Importantly, feature-based attention can even modulate the gain of neurons that do not respond directly to the spatial location of the relevant stimulus, a phenomenon that is thought to globally prime sensitivity to detect relevant features – irrespective of their location – during visual search. However, the neurons in PFC that are thought to provide top-down feedback typically have high-dimensional tuning for multiple features, so it is unclear how feedback selectively modulates responses in neurons tuned to a relevant stimulus without incidentally causing interference by co-modulating neurons tuned to irrelevant features. To investigate this issue, we adapted a spiking neural network model with a first ‘sensory’ layer composed of neurons selective for single features. Neurons in a second ‘control’ layer formed random and reciprocal connections with different sensory neurons, giving rise to PFC-like high-dimensional feature tuning. Stimulating second layer neurons that responded robustly to a relevant stimulus led to corresponding gain modulations in sensory neurons that were directly driven by a relevant stimulus. Importantly, no spurious stimulus-like representations arose in unstimulated sensory neurons – despite high-dimensional tuning in second layer neurons – because the random connections averaged out feedback targeted on irrelevant features. Next, we show that subtly increasing the probability that similarly tuned sensory neurons converge on the same second layer neurons can yield most of the noise-cancelling benefits of completely random connections while simultaneously producing spatially global feature-selective modulations in unstimulated sensory neurons. Collectively, these results suggest that a delicate balance between randomness and structure can support top-down feedback signals that globally enhance sensory neurons tuned to relevant features, without leading to spurious stimulus representations that might interfere with perceptual processing.