artificial neurons that act as microprocessors. Each PE has an input and an output side. The connections are on the input side correspond to the dendrites of the biological original and provide the input from other PEs while the connections on the output side correspond to the axon and transmit the output. Synapses are mimicked by providing connection weights between the various PEs and transfer functions or thresholds within the PEs. Figure 1 illustrates a simple processing element of an ANN with three arbitrary numbers of inputs and outputs [2]. y Output signal to other PEs (corresponds to the axon in biological neurons) y y y = f(T) Transfer function T = Σ (wi xi) Activation function w2 w3 w1 Synaptic weights Inputs coming from other PEs (corresponds to the dendrites of a biological neuron) X2 X3 X1 Figure 1 Processing element of an ANN model with three arbitrary numbers of inputs and outputs The activation of the PE results from the sum of the weighted inputs and can be negative, zero, or positive. This is due to the synaptic weights, which represent excitatory synapses when positive (wi>0) or inhibitory ones when negative (wi<0). The PEs output is computed by applying the transfer function to the activation, which as a result of the synaptic weights, can be negative, zero, or positive. The type of transfer function to be used depends on the type of ANN to be designed. Currently, back-propagation is the most popular, effective and easy to learn model for complex networks [2,3]. To develop a back-propagation neural network, a developer inputs known information, assigns weight to the connections within the network architecture, and runs in the networks repeatedly until the output is satisfactorily accurate. The weighted matrix of interconnections allows the neural networks to learn and remember [4]. In essence, back propagation training adapts a gradient-descent approach of adjusting the ANN weights. During training, an ANN is presented with the data thousands of times (called cycles). After each cycle, the error between the ANN outputs and the actual outputs are propagated backward to adjust the weights in a manner that is mathematically guaranteed to converge [5]. The powder metallurgy processing has the advantage of forming near net shaped components. The advantages of producing complex shapes with close dimensional control at high density (porosity <2%) are 2
RkJQdWJsaXNoZXIy MjM0NDE=