V. Honavar and L. Uhr, "Brain-structured connectionist networks that perceive and learn," Connection Sci., vol. 1, no. 2, pp. 139--159, 1989.  Home/Search   Document Details and Download   Summary   Related Articles   Check

....to this topic) Furthermore, to classify images of dimension wider than 24 24 pixels, as well as to include more recognition capabilities (e.g. translation, rotation, and scaling invariance) the number of neurons and weights greatly grows. The pyramidal system reported by Honavar and Uhr [20] is a specifically tailored multilayer perceptron trained with the backpropagation learning rule, each layer embedded on a pyramidal level. In a layer, a cluster of nodes is connected to a small window in the layer below; the spatial resolution of the layers logarithmically decreases. The learning ....

V. Honavar and L. Uhr, "Brain-structured connectionist networks that perceive and learn," Connection Sci., vol. 1, no. 2, pp. 139--159, 1989.

....They are subject to constant change as implementation technology and strategy improve. 2.3. 1 Performance: speed and efficiency It can be seen from Tables I and III that the basic computing unit in the brain, the neuron, is at least 10 5 times slower then a typical computer switch (see also [55]) In artificial implementations a neural iteration is finished when all (synchronized) nodes in the neural network have calculated their new activation level and possibly updated their weights. In the case of parallel implementations this phase has to be preceded (or partly overlapped) by a ....

Honavar, V., & L. Uhr (1989). Brain-structured Connectionist Networks that Perceive and Learn. Connection Science, 1, 3, 139-159.

....build into the total structure. They have the potential of being made far more brain like, by making the primitive units more neuron like (Shepherd, 1989) and by introducing micro circuits and successively larger assemblies that incorporate more brain like structures and processes (Uhr, 1989; Honavar Uhr, 1990a; 1990b) They are general purpose in the sense of Turing machines, Post productions, and other equivalent universal computers (McCulloch Pitts, 1943; Pollack, 1987) This means that anything that is describable, and therefore can (with enough work to analyze and develop a clear description) ....

....be programmed for a more conventional computer can therefore be handled by some CN. They lend themselves to learning, since their micro modularity means that they can change themselves a little bit at a time, potentially in a way that moves gradually toward adequate behavior (Hinton, 1987a; Honavar Uhr, 1990a) Today s networks can handle only very simple examples of very simple problems. Most have only a few hundred nodes organized into 1 or 2 layer nets. A few much larger networks have been built (chiefly for vision) these are given appropriate larger structures e.g. local connectivity within ....

[Article contains additional citation context not shown here]

Honavar, V., & Uhr, L. (1990a). Brain-Structured Connectionist Networks That Perceive and Learn. Connection Science, 1, 139-159.

No context found.

Honavar, V. and Uhr, L. (1989) Brain-Structured Connectionist Networks that Perceive and Learn, Connection Science 1, pp. 139-159.

.... the successive levels have resolutions of 256x256, 128x128, 64x64, and so on (see below for details) More complex multi resolution representations (see below for details) might be constructed by applying cascades and compounds of transforms to the input image (Uhr, 1972; Hanson Riseman, 1980; Honavar Uhr, 1989). Such representations have been used with considerable success in a variety of image processing applications (Rosenfeld, 1984; Dyer, 1987) Given a suitable set of such transforms, one can construct systems which recognize objects in the input image (Uhr, 1972; Hanson Riseman, 1980; Li Uhr, ....

.... 1984; Dyer, 1987) Given a suitable set of such transforms, one can construct systems which recognize objects in the input image (Uhr, 1972; Hanson Riseman, 1980; Li Uhr, 1987) Generative learning offers a way to discover the necessary set of transforms through feedback guided learning (Honavar Uhr, 1988; 1989). 2. Motivations for Learning from Multi Resolution Representations In spite of the obvious interest and considerable effort in the construction and analysis of multi resolution stimulus encodings as evident from the literature in image processing and computer vision (Rosenfeld, 1984; Tanimoto ....

[Article contains additional citation context not shown here]

Honavar, V., & Uhr, L. (1989). Brain-Structured connectionist networks that perceive and learn. Connection Science - Journal of Neural Computing, Artificial Intelligence and Cognitive Research 1 139-160.

....1973) it is only recently that the need for principled integration of the two paradigms to take advantage of the strengths of both is becoming widely recognized. Recent work by many researchers has resulted in tentative steps in this direction (see for example, Arbib, 1989; 1993; Goldfarb, 1993; Honavar Uhr, 1990b; 1990c; 1993a, 1993b, 1993c; Booker, Riolo, Holland, 1993; Barnden, 1993; Sun, 1993; Dyer, 1993; Honavar, 1992c, 1993; Shavlik, 1993) For some time we have been exploring an approach to learning one that utilizes a range of representations of increasing expressive power which are ....

Honavar, V., & Uhr, L. (1989a). Brain-Structured Connectionist Networks that Perceive and Learn, Connection Science 1, pp. 139-159.

No context found.

Honavar, V. and Uhr, L. (1989a). Brain-Structured Connectionist Networks that Perceive and Learn. Connection Science, 1:139-159. 32 Chapter 12

No context found.

Honavar, V. & Uhr, L. (1989a). Brain--Structured Connectionist Networks that Perceive and Learn. Connection Science, 1:139--159.

....(or, conversely, similarity) between graphs are only beginning to be explored. Another possibility is to investigate restricted subclasses of graphs that lend themselves to efficient distance computation. It is natural to generalize the idea of distance to arbitrary structured representations (Honavar, 1992a; Goldfarb, 1990). Consider a representational framework R (e.g. a space of appropriately defined vectors, strings, trees, graphs) Let I 1 and I 2 be objects represented in R. That is, if R is the space of strings defined over a specified alphabet, say V T , then I 1 ; I 2 2 V T . Key to the notion of ....

Honavar, V. & Uhr, L. (1989a). Brain-Structured Connectionist Networks that Perceive and Learn, Connection Science 1, pp. 139-159.

....approach to learning one that utilizes a range of increasingly complex representations which are incrementally generated as under guidance from the environmental input and feedback. This is a natural extension of generative or constructive learning algorithms for connectionist networks (Honavar Uhr, 1988; 1989; 1990; 1992) for task driven parsimonious generation of CN topologies for tasks such as pattern recognition. 2 Representation Matters A difficulty that plagues both SP as well as CN approaches to machine learning is the choice of inappropriate abstract knowledge representations (e.g. a ....

Honavar, V. & Uhr, L. (1989). Brain-Structured Connectionist Networks that Perceive and Learn, Connection Science 1, pp. 139-159.

Online articles have much greater impact   More about CiteSeer.IST   Add search form to your site   Submit documents   Feedback  

CiteSeer.IST - Copyright Penn State and NEC