Are Cortical Models Really Bound Review by the "Binding Problem"?

of models to perform higher-level visual processing such temporal cortex (IT) described neurons tuned to views of as viewpoint-invariant object recognition in cluttered complex objects such as faces; i.e., the cells discharged scenes has been questioned in recent years by several strongly to a face seen from a specific viewpoint but researchers, who in turn proposed an alternative class very little or not at all to other objects. A key property of models based on the synchronization of large assem- of these cells was their scale and translation invariance, blies of cells, within and across cortical areas. The main i.e., the robustness of their firing to stimulus transformaimplicit argument for this novel and controversial view tions such as changes in size or position in the visual was the assumption that hierarchical models cannot field. deal with the computational requirements of high-level These findings inspired various models of visual obvision and suffer from the so-called “binding problem.” ject recognition such as Fukushima’s Neocognitron Here, we review the present situation and discuss theo- (1980) or, later, Perrett and Oram’s (1993) outline of a retical and experimental evidence showing that the per- model of shape processing and Wallis and Rolls’ VisNet ceived weaknesses of hierarchical models are unsub- (1997), all of which share the basic idea of the visual stantiated. In particular, we show here that recognition system as a feedforward processing hierarchy where of multiple objects in cluttered scenes, arguably among invariance ranges and complexity of preferred features the most difficult tasks in vision, can be done in a hierar- grow as one ascends through the levels. chical feedforward model. Two problems in particular Models of this type prompted von der Malsburg (1981) make object recognition difficult: to formulate the binding problem. His claim was that 1. The segmentation problem. Visual scenes normally visual representations based on spatially invariant feacontain multiple objects. To recognize individual obture detectors (to achieve invariant recognition) were jects, features must be isolated from the surrounding ambiguous: “As generalizations are performed indepenclutter and extracted from the image, and the feature dently for each feature, information about neighborhood set must be parsed so that the different features are relations and relative position, size, and orientation is assigned to the correct object. The latter problem is lost. This lack of information can lead to the inability to