Recognition of Occluded Objects

3 Pattern recognition involves building a mental model to interpret incoming inputs. 4 This incoming information is often incomplete and the mental model must 5 extrapolate to complete the patterns, a process that is constrained by the statistical 6 regularities in nature. Examples of pattern completion involve identification of 7 objects presented under unfavorable luminance or interpretation of speech cor8 rupted by acoustic noise. Pattern completion is also at the heart of other high-level 9 cognitive phenomena including our ability to discern actions from still images or to 10 predict behavioral patterns from observations. 11 Pattern completion constitutes a ubiquitous challenge during natural vision. 12 Stimuli are often partially occluded, or degraded by changes in illumination and 13 contrast. While much progress has been made towards understanding the mecha14 nisms underlying recognition of complete objects, the neural computations under15 lying more challenging recognition problems such as object occlusion remain 16 poorly understood. Here we generically refer to object occlusion to include any 17 transformation where only partial information about an object is accessible (such as 18 the multiple examples illustrated in Fig. 1), and refer to object completion as the 19 ability to infer object identity from partial information (without necessarily 20 implying that subjects perceptually fill in the missing information). 21 Understanding how the neural representations of visual signals are modified 22 when objects are occluded is critical to developing biologically constrained com23 putational models of occluded object recognition and may also shed light on how to 24 solve manifestations of pattern completion in other domains. The development of 25 feed-forward models for visual recognition of whole objects has been driven by 26 behavioral and physiological experiments establishing the hierarchy of feature 27 tuning and robustness to image transformations. Similarly, by systematically 28 examining when and where neural representations that are robust to occlusion