Abstract
Sensory stimuli are represented by the joint activity of large populations of neurons across the mammalian cortex. Information in such responses can be affected by trial-to-trial variability. Because that variability is not independent between neurons, it has the potential to improve or degrade the amount of sensory information in the population response. How potential shared variability at different timescales is structured in neural populations, potential sources of shared variability, and how visual information in specific population is affected by variability all remain open empirical questions. Here, we use neuropixels to simultaneously record tens to hundreds of single neurons in connected structures, primary visual cortex (V1) and lateral geniculate nucleus (dLGN), of awake mice. We measured global and layer-specific shared spike count variability at second to millisecond timescales, and compared shared variability between specific visual drive and a generic electrically-evoked cortical drive. At the population level, we found the shared variability was low-dimensional, and could support a mix of synergistic and redundant coding, depending on timescale. In practice, decoders could use shared variability to enhance fine (<10 ms) information, but shared variability had minimal effect at a timescale of 100 ms in V1. Our results suggest information about visual inputs in shared variability is present, not uniform, and is distributed across physically overlapping but distinct ensembles in V1.NEW & NOTEWORTHY Neurons in early visual areas show variability across repetitions of the same inputs; this variability is shared between neurons nonuniformly within subpopulations of the early visual system, and is nonuniform across timescales. Shared variability in the same responses can be both information-limiting and information-enhancing.