JNTU B.Tech IV ARTIFICIAL NEURAL NETWORKS Previous Question Paper (February/March - 2022 - R18)

A neural network is represented as a directed graph where nodes correspond to neurons and directed edges represent the synaptic connections between them. Each edge is assigned a weight, which quantifies the strength of the connection, while each node performs a specific transformation on its inputs.

The graph structure is typically organized into layers: an input layer, one or more hidden layers, and an output layer. The direction of the edges indicates the flow of information, usually from the input layer toward the output layer in feedforward architectures. Mathematically, the state of a neuron is determined by the weighted sum of the outputs of the neurons connected to it, followed by the application of an activation function. This graph-based abstraction allows for the use of graph theory and linear algebra to model complex data transformations and facilitates the implementation of algorithms like backpropagation, which traverses the graph in reverse to compute gradients.

Error Correction Learning: This technique is based on the principle of minimizing the difference between the actual output of the network and the desired target output. The network iteratively adjusts its weights based on an error signal, defined as the difference between the target and the actual response. A common example is the Delta Rule or the Widrow-Hoff rule, where weight updates are proportional to the error gradient. The goal is to minimize a cost function, such as the Mean Squared Error, driving the network toward a state where the error is minimized over the training set.

Memory-Based Learning: Unlike error correction, memory-based learning relies on storing specific training examples or patterns directly in the network's memory. When a new input is presented, the network retrieves or interpolates from these stored patterns to produce an output. This approach is often seen in associative memories or nearest-neighbor classifiers. Instead of learning a global mapping through weight adjustments, the network effectively 'remembers' the training data, making it highly effective for tasks where local patterns are more important than global generalizations.