Bernal S D_2010.pdf - University of Plymouth

3A. EXISTING MODELS
proposes a trivial extension of the algorithm using the Noisy-OR gate. However, this melhtxl
is only valid for graded variables (see Section 3.3.4 and Pearl (1988)), which is noi the case
for variables in HTM nelworks. .Similarly, the problem of networks with loops is solved by
implementing loopy belief propagation, which is claimed lo provide good results, although no
evidence is provided.
Model simulations shows succesful recognition (72%) of 48 hnc drawing objects (32 x 32 pix­
els) despite translations, distortions and clutter. When tested on the standard Caltech-JOl bench­
mark of natural images, the performance decreased signiHcantly (56%); although when using
their own 4-calegory testset of natural images, the accuracy was very high (92%). Preliminary
results also suggest top-down feedback in the model can account for segmentation, feature bind­
ing, attention and contour completion. Only the last phenomenon is explicitly demonstrated, by
lirstly recognizing a Kanizsa square (input image) as a square (high-level cause), and later alow-
ing lop-down feedback lo increase the response of nodes coding the retinotopic location of the
illusory contours. Due lo the significani similarities between HTMs and the model proposed in
this thesis, a more detailed comparison between them is included in Section A.
3.4.2.2 Model