WBJECA 2025 — Computer PYQ

Explanation:

In a neural network, the weighted sum of inputs at any given node is essentially a linear operation. If we were to stack multiple layers of linear operations without an activation function, the entire network would effectively behave like a single linear regression model, regardless of how many layers it has.

Activation functions (like ReLU, Sigmoid, or Tanh) are applied to introduce non-linearity into the network. This allows the network to learn complex patterns and approximate arbitrary non-linear functions, which is essential for tasks like image recognition, natural language processing, and advanced pattern matching.

We can represent the output $y$ of a neuron with an activation function $f$ as:

$$z=\sum (w_i\cdot x_i)+b$$

$$\text{Output}=f(z)$$

Where $w$ represents weights, $x$ represents inputs, and $b$ is the bias. Without the non-linear function $f$, the network loses its ability to represent complex mapping, effectively collapsing into a simple linear model.