Neural NetworksThe Deep Learning Universe
The network you trained is a feed-forward network: numbers flow one way, from the inputs to the output. Researchers have invented other shapes of network for other kinds of data.
- Convolutional neural networks (CNNs) slide a small window across an image, finding edges first, then shapes, then whole objects. They see for face unlock, medical scans and self-driving cars.
- Recurrent neural networks (RNNs) feed their own output back in with the next input, so they remember what came before. They suit sequences: sentences, music, or a week of temperatures.
- Generative adversarial networks (GANs) set two networks against each other: one forges pictures, the other tries to catch the forgeries, and both get better until the forgeries look real.
- Transformers use attention: every word looks at every other word to decide what matters. They are the networks behind today's chatbots.
Match each architecture with what it does best.
Remembers what came earlier in a sentence
Slides a small window across a photo
A forger network and a detective network compete
Every word pays attention to every other word
Finds edges, then shapes, then objects in an X-ray
Feeds its own output back in at the next step
RNN
CNN
GAN
Transformer
Your journey begins
The field moves fast, and it can look overwhelming. But here is the secret: every one of these giant models is built from what you used in this chapter — weighted sums, activation functions, a loss, and gradients worked out by backpropagation.