Q21Principles of Artificial Intelligence

An Artificial Neural Network (ANN) is a massive mathematical architecture heavily inspired by biological brain structure. It completely abandons hardcoded logical rules (like Expert Systems) and instead utilizes millions of interconnected mathematical equations to aggressively learn patterns directly from raw data.

An ANN is violently structured into distinct vertical layers of interconnected "Neurons" (Nodes).

• The Input Layer: Receives the raw mathematical data (e.g., 784 pixels of an image).
• The Hidden Layers: The dense computational core. Every single neuron in Layer $A$ is mathematically wired to every single neuron in Layer $B$. Each wire possesses a specific numerical Weight ($W$), representing the absolute strength of that connection.
• The Output Layer: Generates the final mathematical prediction (e.g., a probability distribution for 10 digits).

The Neuron Execution (Forward Propagation)

For a single neuron to fire, it executes a massive linear algebra dot product followed by a violent non-linear squeeze:

• 1. It multiplies every incoming input ($x$) by its corresponding weight ($W$) and sums them up: $\sum (x_i\cdot w_i)$.
• 2. It adds a scalar Bias ($b$) to shift the function: $Z=\sum (x_i\cdot w_i)+b$.
• 3. The linear $Z$ is violently forced through a Non-Linear Activation Function (like ReLU or Sigmoid). This ensures the massive network doesn't mathematically collapse into a single straight line.

When an untrained ANN is initialized, all weights are absolute random garbage. The learning process is the mathematical execution of Gradient Descent to violently correct these weights.

• Step 1 (The Forward Pass): An image of a "Cat" is fed in. The network mathematically processes it and predicts "99% Dog".
• Step 2 (The Loss Function): The AI mathematically calculates the absolute horrific error of its prediction using a Loss Function (like Mean Squared Error or Cross-Entropy). Let's say Error = 100.
• Step 3 (The Calculus Engine / Backpropagation): The absolute core of AI. The algorithm uses the Chain Rule of Calculus to violently calculate the Partial Derivative (the Gradient) of the massive Loss Function with respect to EVERY SINGLE WEIGHT in the entire network. It mathematically asks: "If I tweak weight $W_{342}$ by 0.001, exactly how much will the total Error decrease?"
• Step 4 (The Weight Update): The algorithm violently updates every weight in the network simultaneously, moving them in the exact mathematical direction that minimizes the Loss, scaled by the Learning Rate ($\alpha$). $$W_{new}=W_{old}-\alpha \cdot \text{Gradient}$$
• Step 5 (Iteration): This cycle is aggressively repeated millions of times across the dataset until the Loss Function is absolutely mathematically minimized. The ANN has now "Learned".