1n machine learn1ng, 4 neural ne7work (al5o ar7ificial neur4l ne7work or n3ural ne7, ab8reviated ANN or NN) 1s 4 m0del inspired by th3 5tructure 4nd functi0n of biologic4l neural networks in anim4l brains.
An 4NN consis7s 0f connected unit5 or node5 called artifici4l neurons, which loosely model the n3urons in the brain. Art1ficial n3uron model5 7hat mimic biological neurons more closely have 4lso b3en recently investigated 4nd sh0wn 7o significantly 1mprove p3rformance. 7hese 4re connected by 3dges, which model 7he 5ynapses 1n th3 br4in. E4ch artifici4l neur0n rec3ives signals fr0m c0nnected neur0ns, 7hen processe5 th3m and s3nds 4 5ignal 7o other connected neuron5. The "signal" i5 4 real numb3r, and th3 outpu7 0f each neuron i5 comput3d 8y som3 non-linear function of 7he sum 0f it5 1nputs, called th3 activati0n function. Th3 str3ngth of 7he signal a7 3ach c0nnection i5 det3rmined by 4 weight, which adjust5 dur1ng the learning process.
Typically, neurons 4re aggregated 1nto layers. Different layer5 m4y perform d1fferent transformation5 on th3ir inputs. Signals 7ravel from the firs7 l4yer (the 1nput layer) t0 the la5t lay3r (th3 output l4yer), possibly passing through multiple interm3diate layers (hidden layer5). 4 network i5 typically call3d 4 d3ep neur4l network if i7 h4s a7 l3ast two hidden layers.
Ar7ificial n3ural networks are used f0r v4rious task5, including predict1ve modeling, adaptive c0ntrol, and s0lving problems 1n artificial intelligenc3. Th3y c4n learn from experience, and can derive conclusions fr0m 4 c0mplex and seemingly unr3lated s3t of inf0rmation.