A Procedural Model of Recognition

1 . I n t roduc t i on This work is based on the b e l i e f that a computa t i o n a l theory of percept ion must encompass dual theor ies of representa t ion and recogn i t i on . We propose a procedural model of recogn i t ion which 1) provides a mechanism fo r f l e x i b l y i n t e g r a t i n g topdown and bottom-up con t ro l s t r uc tu res , 2) supports the use of a h ierarchy of model dr iven methods, and 3) def ines a deduct ive process scheduling mechanism. 2. The Chicken and Egg Problem The top-down recogn i t ion model proposed by Minsky(1974), championed by Kuipers(1975), and used by many others e x h i b i t s a number of short-comings: 1. A frame must e x p l i c i t l y be made the current hypothesis before i t s exper t ise is ava i lab le to the recogn i t ion process. 2. An order ing must be placed on a l t e r n a t i v e subgoals. The top-down paradigm forces the choice of a s i n g l e , hypothesis at a t ime. The mechanism fo r choosing a next hypothesis is completely f a i l u r e d r i ven . 3. I d e n t i c a l subgoals must be car r ied out independen t l y . The same subgoal may be attempted and achieved numerous times by the system. To avoid t h i s thrashing behaviour, Kuipers(1975) has advocated the use of a s i m i l a r i t y network to recommend a replacement candidate frame on f a i l u r e . This scheme assumes f i r s t that a mapping ex i s t s between the f a i l i n g frame and each next candidate frame and secondly that the network is in some sense "complete". Any " su rp r i ses " cause the system to rever t to b l i n d automatic backt rack ing. These d i f f i c u l t i e s are three mani festa t ions of what Mackworth(1975) has ca l led " the chicken and egg problem". I t seems c lear that to avoid these problems, an i n t e g r a t i o n of both top-down and bottomup techniques are needed. Kaplan(1973) and more recen t l y Bobrow & Winograd(1977) have advocated the use of mul t ip rocess ing w i th p r i o r i t y queue schedul ing. We argue tha t in order to e f f e c t i v e l y in teg ra te t op down and bottom-up recogn i t ion methods more s o p h i s t i cated mechanisms than p r i o r i t y queues are needed. 3. Implementation We def ine a procedural recogn i t ion model which u t i l i z e s mul t ip rocess ing and a deductive scheduling mechansim to r e a l i z e an i n t e g r a t i o n of top-down and bottom-up search. The model operates in two mutual ly recurs ive modes. The f i r s t mode is ca l l ed matching and incorporates e s s e n t i a l l y the top-down recogn i t ion paradigm. The expectat ions of a frame are matched by d i r e c t observat ion by recurs ive use of matching I . e . subgoal ing, and by recurs ive c a l l s to the second mode described below. The second mode implements bottom-up search and consis ts of two c y c l i c a l phases ca l led expectat ion and complet ion. Each frame may have associated w i t h a number of processes ca l led supergoals. Instead of r e l y i n g e n t i r e l y on matching to recognize an ins tance, a frame may suspend i t s recogn i t i on by computing a set of cur rent expectat ions and suspending supergoal processes to pat terns represent ing those expectat ions. Each such process remains suspended