I.;AI{I_Y \"ISI()N AI't'I,I(:A'IIONS OF FEATUI{E-MAI ~ I)IFt"USION-ENttANCEMENT NETS. M. Seibert and A M Waxman I+at,<~rat.ry for %,,nsory t{uholies, (:ollege of Engineering, Boston ITniversity. Boston, MA 02215 USA. We haw. f(,und, using neurological leads, that it is possible to realize the visual learning and recognition of two dimensional ohjects using primardy data-driv~'n processes which achieve invariance to location, size, and orientation of the stimulus. This is accomplished using a spreading activation lwtwork modeled Ks a diffusion r~action over a map of features (e.g., zero-crossings or corners of the initial intensLty map) m o,opcration with a local maxima detection network (an ext.reme kind of competition or contrast enhancement), ;m unsuperwsed learning network using Adaptive Resonance Theory, a complex logarithm mapping, primitive feature detectors, and sinmhm.d saccadic eye ntovelnents. This discussion is focussed on the data-driven early-vision problems to which we applied the spreading act tvat ion (diffusion-enhancement) network. The important role of on-center/off-surround feedback is also briefly considered. Neuroanatomists report that the visual pathway uses extensive local connectivity, simultaneous asynchronous prccessing, and a ",ortical nm~nitication factor" which can be modeled by a complex logarithmic lnappmg. Physiologists report that imagery is quickly proc,ss,.d by speen successful include: • P;uildmg representations which are invariant to ii Indicating the spatio-temporal correspondence of I,+cation, siz,+, and orientation. features in long-range apparent motion. • 1' illinG gaps in contours and completing Extracting features at multiple continuaus I,t~un
ON-CENTER/ OFF SURROUND
~ - -
OFF-CENTER' ON-SURROUND
SHUNRNG FEEDBACK
[
IZED ACTIVITY
T @
ORIENTED OFF-CENTER/ ON.SURROUND
~ _ ~!~.~_#' ~~ FEATURAL T@ @
Im+u
'C,so o c ON-CENTER/ OFF-SURROUND
IMAGE MAP
SURROUND
l:igurel ' l h o N A D E L a c h i e v e s l 0 n g - r a n g e i n t e r a c t i 0 n s u s i n g o n l y local interconnects. Th~ activity maxima are localized and fed hack from the lop level using on-center/off-surround feedback,
Figure 2: Activity clustering over scale. The features (A) are grouped locally' (B). As diffusion continues the global centroid emerges a.q the single activity maximuin(C).
local groupings emerge before global groupings. Thus, attention may be focussed on subparts a.s well as wholes. The convex areas of the act ]vity-space might suggest grouping-regions while the maxima locations might suggest fixation points. The relative heights of the maxima mig~ht further suggest a prioritized scan sequence. The trajectories of the maxima serve to temporally link features. It is striking that so many problems can be approached with ttle same simple underlying mechanism. Spreading activation appropriately influences the activity of distant features, providing a mechanism for long-range interactions between features using only local interc readily implemented in specialized hardware. This abstract describes material report.ed in the following papers. [11 Seib<-rt M., and \¥axman, A.M., Spreading Activation Layers, Visual Saccades, and Invariant Representations for Neural Pattern Hecognition Systems. Laboratory for Sensory Robotics, Boston Unzversily, LSR- TR-5. January, 1988. [2] Waxman. A M., Seibert, M., and Wu+ J., Applications of a Neural Analog Diffusion Layer, Laboratory [or .%nsory Robotics, Boston l'7~ivers21~, in preparation, 1988.
523