What is Deep Learning?

Each layer of a deep network transforms its input into a slightly more abstract representation — early layers in an image model might detect edges, later ones detect shapes, and the final layers detect objects. This layered structure is what lets deep learning scale to messy, high-dimensional inputs like pixels, audio waveforms, and raw text where manually specifying the relevant features would be impractical. Training works by backpropagating an error signal from the output back through every layer, adjusting millions or billions of weights via gradient descent so the network's predictions get closer to the correct answer over many passes through data. Modern large language models are deep networks built almost entirely from a specific layer type — the transformer — stacked dozens of times. Deep learning's strength is pattern generalization from examples; its weakness is that it's opaque (hard to explain why a specific output emerged) and data-hungry, which is why techniques like transfer learning and fine-tuning exist to reuse a pretrained network's knowledge on a narrower task.