Deep learning

Deep learning is a machine-learning approach that trains neural networks with multiple layers to learn representations from data. Its layered computations can build increasingly abstract features.

Artificial neural network: A computational model of interconnected units that transforms inputs into outputs using adjustable parameters. Deep learning uses neural networks with multiple layers as its central computational structure.
Tensor: A multidimensional array used to represent numerical data and parameters in machine learning. Deep-learning inputs, activations, and parameters are commonly stored and processed as tensors.
Convolutional neural network: A neural network that uses convolutional operations to process spatially structured data. Convolutional networks became a prominent deep-learning approach for image recognition.
Perceptron: An early artificial neuron model that classifies inputs using a weighted sum and threshold. It was an influential precursor to the layered neural networks used in deep learning.
Shallow learning: Machine learning with models that use few computational layers or rely on manually designed features. It contrasts with deep learning’s use of multiple layers to learn representations.
Gradient descent: An optimization method that updates parameters in the direction that reduces a differentiable objective. Training uses it to reduce prediction error by changing network parameters.
Linear algebra: The branch of mathematics concerned with vectors, matrices, linear transformations, and systems of linear equations. Neural-network layers rely on matrix operations to transform batches of data.
Recurrent neural network: A neural network whose recurrent connections maintain information across a sequence of inputs. Recurrent networks apply deep learning to sequential data such as speech and text.
Backpropagation: An algorithm that computes how a neural network’s parameters affect its error by propagating derivatives backward through the network. Its adoption made gradient-based training of multilayer networks practical.
Support vector machine: A supervised-learning method that separates classes by optimizing a decision boundary, often with a maximum margin. It is a prominent alternative that can work well without deep neural networks.

Linked from 13 pages