A possible neural net implementation of an index pointer
Koch and Ullman [Ref a] proposed a winner-take-all network that serves as a possible implementation of something like a single visual index – although they view it as an implementation of a mechanism for scanning focal attention. The network below sketches the essential aspect of the network function.
A Winner-take-all network for implementing aspects
of an attention pointer function.
In this illustration, the sensors provide a topographic array of activations in a buffer that holds a retinotopic copy of the sensor activity. This, in turn, feeds into a winner-take-all network that converges on the most active region (let’s call it the focus) of this buffer (where the reason for this peak activity is left open – presumably it is because there is something distinctive or interesting at that location) and turns all other regions off. This can be seen as corresponding to making an index to the focus available, as follows. As a result of the convergence and the inhibition of other areas, it becomes possible to send a “probe” signal through the network that is routed solely to the surviving (noninhibited) focus. This probe can then be used to check whether certain global property detectors fire (assume they were set at just below threshold). If the property-detector for Pi responds, then we know that it is the focus, and not some other region, that has property Pi . Thus we can make property inquiries of the focus of a topographical array. Notice that we are thus able to check on properties of a focal region of the retinotopic display without knowing anything about that region (including where it is!), except that it was the most active region within some area of the visual field. (Other properties of visual indexes assumed in the theory, such as multiplicity of pointers and object-tracking, require additional assumptions (see, Pylyshyn, 1994), but this mechanism at least shows that a pointer is easily implemented in a simple network). (In their paper, Koch and Ullman actually provide a design for a WTA circuit that is guaranteed to converge rapidly on the most active input and to retain the value of that input.)
[a] Koch, C., & Ullman, S. (1985). Shifts in selective visual attention: towards the underlying neural circuitry. Human Neurobiology, 4, 219-227.
[b] Pylyshyn, Z. W., & Eagleson, R. A. (1994). Developing a Network Model of Multiple Visual Indexing (abstract). Investigative Ophthalmology & Visual Science, 35(4), 2007-2007.