In a special issue marking 30 years since the publication of Marr’s Vision (Perception 41:9, 2012), Poggio proposed an update to Marr’s influential “levels of understanding” framework. As well as understanding which algorithms are used for computations such as stereo or object recognition, we also need to understand how observers learn these algorithms, and how this learning is accomplished by neural circuits. I will describe research that addresses this problem in the domain of cue combination. In the last decade, linear cue combination has emerged as a common principle in visual and multisensory processing. In very many tasks, a computational goal (to minimise sensory uncertainty) is achieved by the algorithm of weighted averaging. This framework provides a good description of observers’ behaviour when combining sensory estimates (e.g. multiple depth cues). However, research has repeatedly shown that the computations carried out by developing perceptual systems – up to 8 years or later in humans – are not those leading to uncertainty reduction via weighted averaging. I will describe results showing how developing and mature perceptual systems differ in their computations when combining sensory cues, and outline two key problems for current and future research: 1. understanding the reorganisation of neural information processing that underlies these computational changes, and 2. understanding the learning mechanisms by which we acquire cue combination abilities through perceptual experience.