JNTU B.Tech IV ARTIFICIAL NEURAL NETWORKS AND FUZZY SYSTEMS Previous Question Paper (September 2020 - R16)

Artificial neurons are the fundamental building blocks of neural networks, representing mathematical models of biological neurons. The three primary models are:

1. McCulloch-Pitts (M-P) Model: This is the simplest model, functioning as a binary threshold unit. It receives multiple inputs (x1, x2, ..., xn), multiplies them by fixed weights, sums them, and compares the result against a threshold (θ). If the sum is greater than or equal to θ, the output is 1; otherwise, it is 0. It lacks adaptive learning capabilities as weights are fixed.
2. Perceptron Model: An extension of the M-P model that introduces adaptive weights and a bias. It uses a hard-limiting activation function (step function). The output is determined by the sign of the weighted sum plus bias. It is capable of learning through the Perceptron Learning Rule by adjusting weights based on classification errors.
3. Adaline (Adaptive Linear Neuron) Model: Unlike the Perceptron, Adaline uses a linear activation function during the training phase. It compares the actual output (the weighted sum) to the target value to calculate the error. It employs the Least Mean Square (LMS) or Delta rule to minimize the mean square error, allowing for more stable convergence in linear classification tasks.

Learning laws define how weights are adjusted during the training process to minimize error or optimize performance. The following table summarizes the weight updation rules:

Where η is the learning rate, t is the target output, y is the actual output, and x is the input vector. Hebbian learning follows the principle that neurons firing together wire together, strengthening connections based on correlation. The Perceptron rule adjusts weights only when an error occurs, using the difference between target and actual output. The Delta rule is a gradient descent approach that minimizes the sum of squared errors. Competitive learning updates only the weights of the 'winning' neuron to make it more responsive to the input pattern, facilitating clustering.