JNTU B.Tech III ARTIFICIAL NEURAL NETWORKS Previous Question Paper (November/December - 2020 - R16)

Neural network architectures are classified based on the pattern of connections between neurons and the flow of information. The primary architectures include:

• Single-Layer Feedforward Network: Consists of an input layer and an output layer. Information flows in one direction from input to output, with no feedback loops. It is primarily used for simple pattern classification.
• Multi-Layer Feedforward Network: Includes one or more hidden layers between the input and output layers. This architecture can learn complex, non-linear mappings and is trained using algorithms like backpropagation.
• Recurrent Networks: Characterized by feedback loops where the output of a neuron is fed back as input to itself or other neurons in the same or previous layers. This allows the network to store information about past inputs, making it suitable for temporal sequences and time-series data.
• Competitive Networks: Neurons compete for the right to respond to a given input pattern. Typically, only one output neuron is active at a time, which is useful for clustering and feature extraction.

Learning methods in neural networks are classified based on the nature of the training data and the feedback mechanism provided to the network:

• Supervised Learning: The network is provided with a set of input-output pairs (labeled data). The network adjusts its weights to minimize the error between its predicted output and the target output. Example: Backpropagation algorithm used in Multi-Layer Perceptrons for classification tasks.
• Unsupervised Learning: The network is provided with input data only, without corresponding target outputs. The network discovers hidden structures or patterns in the data by itself. Example: Hebbian learning or competitive learning used in Self-Organizing Maps (SOM) for clustering.
• Reinforcement Learning: The network learns through interaction with an environment. It receives a scalar reward or penalty based on its actions, aiming to maximize the cumulative reward over time. Example: Q-learning used in adaptive control systems and robotics.