Prolonged exposure to an “adapting” visual stimulus improves our ability to detect and discriminate subsequently viewed “test” stimuli that have similar properties to the adaptor. These changes in perceptual sensitivity are often accompanied by reductions in the firing rates and shifts in the tuning curves of sensory neurons. Adaptation may help reduce the redundancy of neural codes, improving how the brain represents prevailing stimulus conditions. It allows a dynamic trade-off between the range of stimuli that can be represented by a neuron’s firing rate and the minimum detectable difference between stimuli. The best functional example of adaptation comes from the retina, where ganglion cells can encode luminance over many orders of magnitude, but adaptation can shift the sensitivity of individual neurons to optimally represent the prevailing conditions. However, the links between neuronal and perceptual adaptation remain poorly understood for even basic stimulus properties such as colour, shape and motion. In particular, while adaptation is well characterised at the level of single neurons, few studies have examined how adaptation affects the responses of populations of neurons. This is important, because perception depends on both the firing rates, and structure of response correlations, across a population of neurons. To better understand how adaptation affects sensory encoding by neuronal populations, we implanted multi-electrode arrays in a range of visual cortical areas in anaesthetised marmosets. The marmoset is an ideal model for array recordings because its smooth cortical surface ensures most brain areas are readily accessible with electrode arrays. We recorded visual-evoked spiking activity at up to 128 sites in V1, V2 and the dorso-medial area (DM). We used two main stimulus sets: first, direction tuning was characterised in an unadapted case using moving gratings; second, direction tuning was characterised after adaptation to 1-2 seconds of sustained exposure to motion in a single direction. In addition, we characterised the receptive field locations of all neurons. At the single neuron level, the effects of adaptation depended on the cortical area. In all areas, neuronal gain, or the peak firing rate evoked by a test stimulus, was reduced following adaptation, regardless of the relationship between a neuron’s preferred direction and the adaptation direction. In V1/V2, adaptation produced “repulsive” changes in tuning, such that the post-adaptation preferred direction was shifted away from the adaptation direction. Surprisingly, in DM, no consistent changes in direction tuning were observed following adaptation. To characterise effects at the population level, we determined spike-count correlations (rSC). For each pair of neurons, rSC is the Pearson correlation between the spike counts evoked by multiple repetitions of an identical stimulus. Before adaptation, rSC between a pair of neurons was highest when the neurons had similar preferred directions and had the greatest level of overlap in their receptive fields. This is expected from previous work, and reflects the larger proportion of inputs shared between neurons with similar sensitivity. After adaptation, average rSC across the population remained constant, but the structure of correlations changed in a tuning-dependent manner. rSC decreased for neurons with similar direction preferences, leading to a smaller range of correlations across the population. We argue that this decorrelation is a way of optimising sensitivity to the prevailing stimulus conditions, consistent with Barlow’s original hypothesis that adaptation decorrelates the activity of a neuronal population to reduce redundancy. However, the tuning-dependence of the decorrelation suggests that adaptation is only able to affect a limited range of inputs to a population of neurons, notably those that are shared locally, between neurons with similar preferences, rather than those that are globally shared across all neurons in an area. The implications of this tuning-dependent correlation for perception are captured in a population decoding model.