SOME COMMENTS ON INTHlACTIVE AIDS DJ

STATISTICAL SIGiVL ANALYSIS AND PATIEEN RECOGNITION

Charlton M. Walter Data Sciences Laboratory

A i r Force Cambridge Research Laboratories Bedford, Massachusetts

Synopsis

The wide gap between the ac tua l performance of a variety of automatic "target" detection and categorization devices, and the anticipated p e r f o m c e of the devices, based on seemingly reasonable mdela of the processes involved, led t o our u t i l i za t ion , i n t he ea r ly lSO's, of a dedicated medium size processor as an a id in exploring the origin of sore of t h e s t a t i s t i c a l difficult ies. This system, consisting of a pa i r of interconnected PDP-1 computers, coupled t o both color and blacklwhite displays, permitted the visual izat ion of s ignal data f rom real sources , t rea ted as ensembles of points and as t r a j ec to r i e s , i n su i t ab le N-dimensional spaces. The e f f ec t of var ious da ta f i l t e r ing and transformation operations could be observed by projecting both raw and t r a n s f o m d d a t a on planes specified by sui tably selected pairs of vectors in the space. Color coding was used f o r a variety of purposes: dist inguishing ensembles of data from di f fe ren t sources, or to indicate t h e from some epoch, o r degree of correlation between d i f fe ren t po in ts .

We have found this concrete ar t iculat ion of vector space, coupled with close man-machine interaction, to be of p a r t i c u l a r l y p e a t v a l u e i n obtaining insight into why s o many pa t te rn categorization schems have worked so wel l when few categories-and only mdest size samples-were involved,, and f a i l ed r a the r s ign i f i can t ly under mre r ea l i s t i c cond i t ions .

To the researcher , th i s in te rac t ive ab i l i ty has proved of value i n narrowing the domain of reasonable hypotheses, through the construction of counter examples. However, t h i s added ins ight is often obtained at the price of considerable f rus t ra t ion in no t having the f lex ib i l i ty of communication with the processor needed to quickly synthesize new modes of action. Additional f rus t r a t ion is incurred upon finding that the requested action took too long to execute . To the systems engineer, the abil i ty to quickly see the e f f ec t s of perturbations-and of approxinrition-on a systems design, has been of major importance.

The a r t i cu la t ion of signals and signatures as points and t r a j e c t o r i e s i n N-space is also proving of great value in enlightening non-technical people on SOIE of the nuances of s t a t i s t i c a l signal analysis and in t ak ing some of the aura of mystery out of the complex processes which a re involved. This aspect of an interactive system is of considerable importance in the context of govermnt agencies which have mandates fo r

overseeing the development of complex and very expensive systems, whose performance is of ten l imited by s t a t i s t i ca l p rocesses which are not e a s i l y understood by the layman.

The capacity of an interact ive color display for present ing large amounts of s t a t i s t i c a l information, coupled with the data transformtion requirements needed to adequately visual ize various signal analyeis procedures, have put such a g rea t s t r a in on the central processing system that it i s no longer a matter of attempt- ing to t ime share one processor among many users, but ra ther a problem of how many micro-processors (both l is t -or iented and arithmetic-oriented) can be dedicated t o t h e given systems investigation session. We have found time sharing t o be feasible only to the extent of permitt ing compatible users simultaneous access t o the system. On-line debugging operations can usually be carried on concurrent with an in te rac t ive systems investigation. And, when considerable in te rva ls of time are to be spent contemplating a s ta t ic d i sp lay , it has been possible to car ry on low leve l in te rac t ion f rom several display stations. Generally, however, t h e l a t t e r operation has prayed t o be ra ther f rus t ra t ing , since additional constraints had t o be imposed on the modes of operation-of an executive system which is s t i l l too l imi ted to be r e a l l y sa t i s f ac to ry

The poten t ia l of interactive systems is very great, but clearly hinges on having an extremely flexible executive system. This requires substantial memry and list processing capability. In addition, such frequent use is made of inner- product operations that a special inner-product forming arithmetic unit is extremely desirable. Fortunately, recent advances in digital systems technology hold promise, within the next five years, of making processor/disphy systems, t a i l o r e d t o t h e somewhat unusual demands of t h i s type of interactive system, both economically feas ib le and highly pract ical .

The cost of dedicated processor capability needed t o e f f e c t i v e l y implement some of these interactive techniques for evaluating and analyzing pattern recognition schemes has been high, but was negligible compared t o t h e c o s t that has gone into the hardware implementation of a Large number of abortive pattern recognition devices.

44-1