Backpropagation (தவறுகளை திருத்துதல் / Learning from Mistakes)

Let's complain to the Hotel Manager! (ஹோட்டல் மேனேஜர்ட்ட கம்ப்ளைன்ட் பண்ணுவோமா!)

Technical Meaning: பின்னோக்கு பரவல் (Pinnokku Paraval) - Sending errors backward to adjust weights and learn.

The Core Idea

Backpropagation (Backward Propagation of Errors) is how a neural network actually learns. After Forward Propagation makes a guess, the network checks how wrong the guess was. It then sends that "error" backward through the network, layer by layer, adjusting the weights to make sure the mistake is smaller next time.

The Origin Story

Before the 1980s, neural networks were painfully slow to train because no one knew an efficient mathematical way to update the weights in deep layers. Geoffrey Hinton and others popularized the backpropagation algorithm, which uses calculus (the Chain Rule) to efficiently calculate exactly how much every single weight contributed to the final error, revolutionizing modern AI.

The Tamil Analogy

Continuing our Saravana Bhavan (சரவண பவன்) analogy:

Suppose the waiter served you a Masala Dosa, but it was completely burned! You complain to the manager. This is the Error.

The manager doesn't just yell at everyone randomly. He traces the error backwards (Backpropagation):

1. First, he goes to the delivery boy: "Did you drop it?" (Delivery boy is fine).
2. Then, he goes backwards to the Dosa Master: "Did you burn it?" The Master says, "Yes, the pan was too hot." (Manager adjusts the Master's 'weight' by warning him).
3. Then, he goes further backwards to the supervisor: "Why didn't you check the temperature?" (Adjusts supervisor's 'weight').

The blame (error) is distributed backwards to everyone who caused it, so the next time an order is placed, the Dosa comes out perfectly!

Try It Yourself