Foundations
Deep Learning
Also known as: deep neural networks
The branch of machine learning that uses many-layered neural networks to learn complex patterns directly from raw data.
Deep learning is the technique behind almost every recent AI breakthrough, from photo recognition to chatbots. It uses artificial "neural networks" with many layers stacked on top of each other. Each layer learns to spot slightly more complex patterns than the one before, so from raw pixels or raw text the system gradually builds up an understanding, without a human spelling out the rules.
Deep learning is the sub-field of machine learning that uses neural networks with many layers, hence "deep". Its defining strength is representation learning: rather than engineers hand-crafting the features a model should look at, the network learns useful features by itself, layer by layer, from raw data.
Why it took off
Three things came together in the 2010s: much larger datasets, far more powerful hardware (GPUs), and better training methods. That combination let deep networks outperform older approaches on images, speech and language.
Where it leads
Deep learning produced the transformer, which in turn made today's large language models and generative AI possible. When people say "AI" today, they almost always mean a deep-learning system.
Deep learning refers to hierarchical, differentiable models, primarily multi-layer neural networks, trained end-to-end by gradient descent with backpropagation (LeCun, Bengio & Hinton 2015). Its central premise is learned representation: composing simple non-linear transformations into deep stacks yields features of increasing abstraction, obviating manual feature engineering and enabling strong performance on high-dimensional, unstructured inputs.
The modern era rests on scale, larger models, datasets and compute, and on architectural advances (convolutional networks for vision, then the transformer for sequences) plus enabling techniques (ReLU activations, normalisation, residual connections, Adam optimisation, dropout). Deep learning underpins essentially all frontier AI, including LLMs and diffusion models. Open questions concern data efficiency, robustness and interpretability: deep networks are powerful function approximators whose internal decision processes remain only partially understood, motivating the fields of explainability and mechanistic interpretability.