Questions
3 questions per university semester paper
Difficulty
Medium
Importance
High yield for core Data Science and AI modules
Overview
Neural Networks are computational models inspired by the structure of biological neurons, serving as the foundational building blocks for modern Deep Learning. Understanding these concepts is critical as they form the bedrock of AI-based applications in university curriculum and technical interviews. Aspirants must focus on the mechanism of information processing, gradient optimization, and architecture-specific utility.
Perceptron and Activation Functions
A Perceptron is the simplest form of a neural network, consisting of a single neuron that maps input vectors to binary outputs using weights and a bias. Activation functions are mathematically necessary to introduce non-linearity, enabling the model to learn complex, non-linear data patterns.
- Weighted sum calculation: z = Σ(wi * xi) + b
- Threshold logic: y = 1 if z > 0 else 0
- Sigmoid function range: (0, 1)
- ReLU (Rectified Linear Unit): f(x) = max(0, x)
- Hyperbolic Tangent (tanh) range: (-1, 1)
Backpropagation
Backpropagation is the primary algorithm used for training artificial neural networks through the method of gradient descent. It involves calculating the gradient of the loss function with respect to each weight by applying the chain rule of calculus, moving backward from the output layer to the input layer.
- Objective: Minimize total error (Cost Function)
- Algorithm type: Supervised Learning
- Uses Chain Rule for partial derivatives
- Gradient Descent: Weights updated by -learning_rate * gradient
- Requires a differentiable loss function
CNN and RNN Basics
Convolutional Neural Networks (CNNs) are specialized architectures for grid-like data such as images, utilizing filters to extract spatial features. Recurrent Neural Networks (RNNs) are designed for sequential data, maintaining an internal state (memory) to process time-series inputs or text.
- CNN key operations: Convolution, Pooling, Stride
- CNN utility: Image classification and Computer Vision
- RNN structure: Looped architecture for temporal data
- RNN issue: Vanishing gradient problem
- RNN application: Natural Language Processing (NLP)
Formula Sheet
z = Σ(wi * xi) + b
σ(z) = 1 / (1 + e^-z)
w_new = w_old - η * (∂Loss / ∂w)
Exam Tip
Always draw a labeled diagram of a single neuron when explaining the Perceptron to secure full marks on descriptive theory questions.
Common Mistakes
- Confusing the role of bias with weights in the perceptron formula.
- Assuming backpropagation updates weights in the forward pass instead of the backward pass.
- Incorrectly stating that ReLU is bounded between 0 and 1.
More Revision Notes
Ready to test yourself?
Play topic-wise Neural Networks Basics questions in Aspirant Arcade — gamified MCQ practice.
Download Free