When an image is presented to one eye and a very different image is presented to the corresponding location of the other eye, the two images compete for conscious representations, such that only one image is visible at a time while the other is suppressed. Called binocular rivalry, this phenomenon and its deviants have been extensively exploited to study the mechanism and neural correlates of consciousness. In this paper, we propose a framework - the unconscious binding hypothesis - to distinguish unconscious processing from conscious processing. According to this framework, the unconscious mind not only encodes individual features but also temporally binds distributed features to give rise to cortical representations; unlike conscious binding, however, unconscious binding is fragile. Under this framework, we review evidence from psychophysical and neuroimaging studies and come to two important conclusions. First, processing of invisible features depends on the "level" of the features as defined by their neural mechanisms. For low-level simple features, prolonged exposure to visual patterns (e.g. tilt) and simple translational motion can alter the appearance of subsequent visible features (i.e. adaptation). For invisible high-level features, complex spiral motion cannot produce adaptation, nor can objects/words enhance subsequent processing of related stimuli (i.e. priming). Yet images of tools can activate the dorsal pathway. Second, processing of invisible features has functional significance. Although invisible central cues cannot orient attention, invisible erotic pictures in the periphery can nevertheless guide attention, likely through emotional arousal; reciprocally, the processing of invisible information can be modulated by attention.