Strong and Weak Connections
Have you ever practiced something so many times that you got really good at it? Maybe you practiced kicking a soccer ball, or reading a tricky word, until it felt easy. Your brain actually changed when you practiced — the connections between certain neurons got stronger! Neural networks have something very similar going on inside them.
What Is a Weight?
Every connection between two artificial neurons has a number attached to it. That number is called a weight. A high weight means the connection is strong — signals travel through it powerfully. A low weight means the connection is weak — signals barely get through. Think of weights like the volume knob on a speaker. Turn it up high and the music blares loudly. Turn it down low and you can barely hear a sound. Weights do the same thing for signals: they turn the signal up or down before it moves to the next neuron.
A weight is a number that controls how strong a connection is between two neurons. Strong connections (high weights) let signals through powerfully. Weak connections (low weights) barely let signals through.
Here is a story. Imagine you have two friends, and both of them tell you the weather before school. One friend always gets it right. The other friend is wrong almost every day. Pretty soon you trust the first friend much more. In a neural network, weights work the same way. If a connection reliably helps the network get the right answer, its weight grows higher. If a connection leads the network in the wrong direction, its weight stays low or gets smaller. The network learns by adjusting weights — turning some up, turning others down.
Complete the sentence about weights.
A neural network can have millions of connections, each with its own weight. Before the network has learned anything, all those weights start as random numbers — the network is just guessing. But as it practices, it adjusts those weights little by little, getting better and better at its job. When a network is fully trained, its weights hold everything it has learned. Change the weights and you change what the network knows!
A neural network does not store memories the way you do — with words and pictures in your head. It stores everything as numbers (weights). There is no little picture of a cat saved anywhere. The cat-knowledge is spread across millions of weight numbers all at once.
What is a weight in a neural network?
What happens to weights as a neural network practices?
Weight It Out
- Find two dice (or write numbers 1-6 on small pieces of paper and draw one randomly).
- You and a partner will each roll one die. Your number is your 'weight'.
- The player with the higher weight speaks loudly and clearly. The player with the lower weight must whisper very softly.
- Take turns rolling and repeating a fun fact — the high-weight player speaks it loudly, the low-weight player whispers it.
- Notice how the high-weight message is easier to hear. That is exactly what strong connections do for signals in a neural network: they make the signal louder and more important.