How much does each taste count?

Four numbers arrive from the bottle. A price has to come out. Something in between has to decide how much each of those four numbers counts.

Think about what that means for a moment. Apricot might turn out to be a strong sign of an expensive wine. Or it might mean nothing at all. The model has no idea which, so it needs a way to say how much this particular taste should be trusted when working out a price.

That is what sits on each connection: a number saying how much the price should trust what that unit reports. It is the connection strength from earlier, now given an actual value.

The number has a name. It is called a weight.

So the connection is the wire between a unit and the price, and the weight is the number sitting on that wire. The wire never moves. The number does.

A high weight on the apricotty connection means apricot pushes the price up strongly. A low weight means apricot barely matters. Same wine either way. The difference is entirely in the weight.

Weights are not special to wine. Any time a neuron takes in several signals, each one arrives through a connection with a weight, and the weight decides how much that signal counts. Module 04 builds this idea from scratch with a different example: connections can count differently.

It helps to picture each weight as a dial you can turn. Turning a dial up makes that taste count for more. Turning it down makes it count for less.

Now, where do the four dials start?

All four start at 8.

Why 8?

There is no wine-related reason at all. The 8 was not learned from bottles, calculated from prices, or chosen because anyone believed each taste was worth eight dollars.

We picked it before training because it makes the first calculation easy to follow.

We could have started at another number. The model's first guesses and its path through training would be different, but that starting number would still contain no knowledge about wine.

So 8 is a placeholder. It tells us only where the four dials happen to be before the model has seen any evidence.

All four begin at the same placeholder for another reason: no input starts with more influence than another. Cabernet, viognier, plum, and apricot all get the same weight. Not because the model knows they are equally important, but because it has no evidence yet that they should be different.

The model therefore begins with an arbitrary opinion, not a blank. Training is the process that replaces that arbitrary starting point with weights shaped by bottles and prices.

Every useful difference between the four weights at the end will be something learning put there.