Published on
How Do Neural Networks Work? Neurons, Layers and Weights Explained — AI Tutorial, Episode 2

Inside an artificial neural network: neurons, layers and weights
In episode 1 we saw that artificial neural networks (ANN) are the hot topic in AI — powerful, and usable across a remarkably broad set of domains. Episode 2 takes the lid off. movingimage CEO Dr Ingo Hofacker walks through what an ANN actually looks like and how a signal travels through it, step by step and without a single formula you need to memorise.
A mathematical model of the brain
An ANN is, in essence, a mathematical model of the human brain. For our purposes it is enough to remember that the brain is made of trillions of neurons connected to one another through synapses. When the brain is active, neurons talk to each other by sending current through those synapses to their neighbours. A few rays of light touch the neurons on your retina, electrical signals travel onwards until they reach the visual cortex — and bingo, the signal is classified as coming from your cat.
Layers: why they are called deep neural networks
Now build one. You start with neurons, and each ANN consists of a large number of layers of them — which is exactly why they are often called deep neural networks. Every neuron on one layer connects to some of the neurons on the next layer. But not all connections are equal: like in real life, some are stronger than others. That strength is called the weight, and depending on it the signal passing along the connection is either amplified or subdued.
How a signal travels from input to output
The initial signal — a vector of numbers — hits the input layer, the retina in our example. As it travels through the deeper layers it is constantly altered, until it eventually reaches the final layer, also called the output layer. Each time the vector passes from one layer to the next, it is modified by a multiplication with the weights in between those two layers. That is the whole mechanism.
Which raises the obvious question: this is not learning. The knowledge is already there, represented in the weights. So how do we arrive at those weights in the first place — how is a neural network actually trained? That is exactly where episode 3 picks up. Watch episode 2 above, then keep the momentum and move on to the next one.


.avif)


.avif)




.avif)





