Deep learning compared with classical computation.
(a) Classical computer programs convert an input (e.g. noisy image of a cell) into a desired output (e.g. sharp image) via an algorithm with known rules and parameters (‘known routine’). On the other hand, NNs are trained with paired of corrupted and ground-truth images, e.g. a noisy and its equivalent high-quality image of a cell. During training (b-i), the untrained network (dark grey) learns to transform the inputs (left) into the output (right) by observing a large number of paired examples from the training dataset. After training (b-ii), the trained network (light grey) can be used to perform the task similarly to a conventional algorithm on novel data, therefore providing the output from new input data. The large black arrowheads represent dataflow.