SmallAI is a feed-forward neural network with 87 brain cells (each cell is
a number it can change). It has an input layer (where every picture it
knows comes in), a hidden layer of just 3 neurons, and an output layer
that votes for the answer. Because it has far more numbers than TinyAI, it can
store many pairs — each pair is a path through the network.
The dots. The larger dots are the neurons (the places where
the math happens). The smaller dots scattered between them are the
weights — one per wire — and the dots tucked beside each neuron are the
biases. Together those weights and biases are the 87 numbers the brain
can change.
The wires. Each wire is a connection that carries a signal from one
neuron to the next. The thicker and brighter a wire, the more strongly it
connects two ideas. When you teach SmallAI a pair and press Learn, it
strengthens the wires along the right path and weakens the wrong ones.
This is a classic multilayer perceptron (MLP) with one hidden layer. It
can't perfectly separate every possible set of pairs, but for the everyday
picture pairs on this page it usually finds a good set of paths — and you can
watch its confidence bar grow as the right answer lights up.