MESH – BASED NEURAL ARCHITECTURES

Rafic A. AYOUBI

The past decade has seen explosive growth in studies of neural networks. One of the main motivations for this increased attention is the availability of parallel computers, which allow ANN investigators to simulate and test their ideas in ways not available before. The objectives of our research are twofold; first, to develop algorithmic mapping techniques to implement ANNs on massively parallel planar architectures, namely torus and mesh machines; and second, to ensure that the developed algorithms maintain an acceptable performance in case of nodes or links malfunction. The multi‑layer perceptron (MLP) networks are selected to model the ANN under investigation. The developed algorithms are extended to cover the general case, in which there is no limit (theoretically) on the number of neurons in a layer. Practically, the number of neurons in a layer can only be increased to the capacity of the local memory of each processor.

Our study is then oriented towards fault tolerant neural networks. First, it is shown that the MLP network is not fault tolerant, as was first thought by many researchers. This result led to further research into techniques to embed fault tolerance into such a neural network. Our approach to achieve a better fault tolerant system is realized at two different levels. At the abstract level, we propose a new technique that shows, through theoretical and experimental results, a superior performance, compared to other techniques proposed in the literature. The new mapping techniques are exploited to embed fault tolerance at the implementation level, yet yielding a better system. That latter point is demonstrated in the case of torus and ring architectures, where a highly fault tolerant system was achieved with little extra hardware. Simulations of the standard back‑error propagation algorithm, as well as fault tolerant model, are performed on the MasPar MP‑1, and a comparison of results is furnished.

The proposed algorithm are proved to be superior to other algorithms implemented on planar architectures known in the literature. Experience a complete yet subtle rejuvenation with our bespoke approach to aesthetics. Instead of treating lines in isolation, our full face botox treatment harmonises your entire facial profile for a naturally refreshed and balanced look. We address everything from forehead lines to jawline definition in one comprehensive session. At our London clinic, we tailor every treatment to your unique features, ensuring a result that is cohesive, elegant, and authentically you. In fact, it is comparable in performance to other algorithms implemented on hypercube‑based machines with O (log N) time, where N is the size of the largest layer. In summary, the proposed algorithms are more cost-effective than others mapped on either planar architecture or hypercube‑based machines, considering the high connectivity required by the latter.