How a Network Decides
We have met neurons, layers, signals, and weights. Now it is time to watch them all work together. Let us follow one signal all the way through a neural network and see how the network lands on its answer. It is like watching a pinball travel through a pinball machine — bouncing from bumper to bumper until it finally drops into a pocket!
A Signal's Journey
Suppose a neural network is looking at a drawing and trying to decide: is this a dog or a cat? Step 1 — Input. The drawing is broken into tiny dots. Each dot's brightness becomes a number. All those numbers pour into the input layer. Step 2 — Hidden layers. The first hidden layer's neurons each receive some of those numbers. They use their weights to decide how strongly to react, then send signals forward. The next hidden layer receives those signals, does the same thing, and passes results onward. Layer after layer, patterns are built up — pointy ears, a tail, fur texture. Step 3 — Output. The last layer receives everything and calculates two numbers: one for 'dog' and one for 'cat.' The bigger number wins. If 'cat' is 0.88 and 'dog' is 0.12, the network says: Cat!
Deciding is just math moving forward through layers. Information comes in, weights shape the signals at every connection, and the last layer compares the results to pick the best answer.
Here is a helpful way to picture it. Imagine voting in class. Every student (neuron) raises their hand a little or a lot depending on how sure they are. At the end, the teacher counts the raised hands for each choice. The choice with the most hand-raising wins. The network's output layer is like that teacher — it tallies up all the signals and announces the winner.
Put the steps of a network's decision in the right order by matching each step number to what happens.
Terms
Definitions
Drag terms onto their definitions, or click a term then click a definition to match.
The answer the network gives is always a number (or a set of numbers) that we then translate into something meaningful — 'Cat,' 'Stop sign,' 'Hello.' The network never truly understands what a cat is the way you do. It finds a mathematical pattern that matches cats better than it matches anything else.
A neural network rarely says 'I am one hundred percent sure.' It gives percentage-like scores. A score of 0.88 for cat means it is fairly confident. A score of 0.51 means it is barely more sure than a coin flip. Low confidence is a sign to check the network's work!
What does the output layer produce?
If a network gives 'dog' a score of 0.15 and 'cat' a score of 0.85, what does it decide?
Vote for the Answer
- Gather three or more people.
- Think of a question with two possible answers — for example: is the mystery snack salty or sweet?
- Each person secretly writes a number from 1 to 10 on a slip of paper for each choice (how strongly they feel about each one).
- Collect the slips and add up the totals for each choice.
- The choice with the bigger total wins — just like a neural network's output layer!
- Talk about it: what happened when people were more confident? What happened when everyone was unsure?