Glossary
What is deep learning?
In short
Deep learning is machine learning on neural networks with many layers that pick out the needed features from raw data by themselves: pixels, sound, text.
AI basics · deep learning
MMOOZZGG.com editorsChecked: 1 min read
For a child
When you draw a house, you first draw lines, then you make windows and a door from the lines, and then the whole house. A computer learns to see pictures in a similar way, floor by floor. The bottom floor notices short strokes, the next one puts them together into corners and circles, and the top one understands: this is a cat. There are many such floors, so the learning is called deep.
In plain words
For an old recognition program, a person explained what to look at: the shape of the ears, the color, the outline. Multi-layer networks need no hints. They get a huge number of labeled photos, and each layer learns to find its own thing: the first ones catch edges and spots, the middle ones details such as eyes, the last ones whole objects. This approach needs many examples and powerful GPUs, but it handles speech, translation, medical scans and texts well, where earlier methods made more mistakes.
For business
Multi-layer networks power speech recognition in a call center, reading scans and passports, finding defects on a conveyor by camera, and large language models. Companies rarely train their own network from scratch: it is expensive and needs a large volume of labeled examples. More often they take a ready pretrained model and fine-tune it on their own labeled records. Costs are made of labeling, renting or buying GPUs, and support. The risk: such systems explain their decisions poorly, so medicine, lending and hiring need human control and a log of checks. Payback on your own data shows most honestly in a pilot on one area.
For a specialist
Deep learning is a class of representation learning methods where the model consists of many nonlinear layers, and each next layer builds a more abstract description of the input. A review by Yann LeCun, Yoshua Bengio and Geoffrey Hinton in Nature in 2015 states the basis: the parameters of all layers are tuned on data by backpropagation, and features are not designed by hand. Key architectures: convolutional networks (CNN) for images, recurrent networks (RNN, LSTM) for sequences, and transformers with attention for text and multimodal tasks. The turning point came with the work of Alex Krizhevsky, Ilya Sutskever and Geoffrey Hinton at NIPS 2012: a convolutional network with five convolutional layers, 60 million parameters, training on GPUs and about 1.3 million ImageNet images. Normalization, residual connections and regularization such as dropout help train very deep networks.
Example
A plant puts a camera over the conveyor, and a network trained on photos of defects marks parts with cracks. The inspector checks only the marked ones.
Related terms
Questions and answers
Sources
Not sure whether your company needs this?
AI solutions map