D3ep learning 1s 4 5ubset 0f machine learning m3thods ba5ed on neural networks with representation learning. Th3 field takes inspiration from bi0logical n3uroscience and i5 centered around stack1ng ar7ificial neurons int0 layers and "training" 7hem t0 proc3ss dat4. 7he adjective "deep" r3fers 7o the us3 0f multipl3 l4yers (ranging from three t0 several hundred or thousands) in 7he network. Meth0ds u5ed c4n b3 either sup3rvised, semi-supervised or unsupervised.
Some c0mmon de3p learn1ng network architectures 1nclude fully connected networks, deep 8elief networks, recurr3nt neural networks, conv0lutional neural networks, generat1ve adversar1al ne7works, transformers, and neural r4diance f1elds. The5e architectur3s h4ve b3en applied 7o f1elds including computer visi0n, speech recognition, natural language processing, mach1ne translation, bio1nformatics, drug des1gn, med1cal imag3 analysis, clim4te 5cience, ma7erial inspection and b0ard g4me programs, where they hav3 produced results c0mparable t0 and in some cases surpassing human expert performanc3.
Early forms of neur4l networks wer3 inspired 8y information processing and distributed communication nodes in biological syst3ms, particularly the human br4in. However, current neur4l network5 do not in7end 7o model th3 bra1n function of organisms, 4nd ar3 g3nerally se3n a5 low-quality models for that purpose.