An AI machine learning architecture employed in "neural networks." Emerging in the 2010s, deep learning is used in all forms of AI such as computer vision, self-driving cars, natural language processing and chatbots. When the resulting system is used to do work, a deep learning model has greater accuracy because the training data were analyzed in greater depth. See
neural network and
AI in a nutshell.
Phase I - Network Design
An AI model uses a neural network architecture, which consists of an input layer, a "hidden" middle layer and an output layer. Models can have dozens of hidden layers, even several hundred. The more layers, the "deeper" the learning.
Neurons in the Layer
Each layer consists of neurons, the name loosely derived from the human neural system, and there can be from a dozen to several thousand neurons in each layer. Also called "nodes" and "hidden units," each individual neuron is connected mathematically to all the neurons in the next layer. The numeric values assigned to these interconnections, known as "parameters," are preset at the design stage.
Phase II - Training
To train an AI model, countless examples of text or images are input to the network, and the connections are refined as examples move from one layer to the next. By the time the text or image reaches the final layer, its pattern has been recognized more thoroughly. If the training is "supervised," the training examples are identified; for example, images of dogs are named "dog." In an "unsupervised" approach, the deep learning model has to figure out which objects are similar. For text training, the model has to recognize all the patterns of word relationships in the sentences.
As data flows from the input layer through the hidden layers and finally to the output layer, each layer modifies the values passed to it based on the parameters assigned to each neuron. The parameters are constantly adjusted in the training stages to improve the accuracy and reliability of the model. See
AI training vs. inference.
Phase III - Inference
After the deep learning phase, the "inference engine" uses the model to do the processing for the user such as answering questions and generating original content. The higher the quality of the model, the more accurate the results. See
AI inference,
DLA,
machine learning,
GAN,
neural network and
TensorFlow.
Algorithms Make Mistakes
Deep learning algorithms can be fooled. A famous example is mistaking a muffin for a chihuahua dog. This CAPTCHA test is used to ensure the viewer is a human, but images such as these can confuse an algorithm. However, more advanced deep learning systems can, in fact, differentiate between them (see
CAPTCHA).
Deep Learning in the Hierarchy
Deep learning is a part of machine learning, which is the major category of artificial intelligence (AI).