Created Friday 15 November 2013
A neural network is basically a set of neurons working together in sequence and in parallel. The simplest neural network consists of a single neuron.
The diagram above shows a neural network made up of a single neuron with 3 inputs and 1 output. In its most basic form, each input is multiplied by a neuron weight. The output is the sum of these products passed through an activation function. See the neuron page for details.
The next level of complexity is to have neurons in a layer work together in parallel.
The diagram above shows a single layer neural network with 2 neurons, 3 inputs, and 2 outputs.
The next level of complexity is to have layers work together in sequence.
This diagram shows a multi-layer neural network with:
- 3 inputs
- 1 output
- An input layer with 3 neurons
- An output layer with 1 neuron
- it is fully connected because each neuron is connected to all inputs in the previous layer
Formal Definition
Neural networks can be thought of as a function of the input and the weights of the neurons. More importantly, if each neuron represents a continuous and differentiable function, then the entire network is a continuous and differentiable function. This property is important for learning. (See +BackPropagation).
Neural Network Families
Unfortunately, a "universal" neural network suitable for any type of problem does not exist. There are actually different types of NNs suitable for different types of problems:
- recursive neural networks - sentiment analysis (natural language processing)
- convolution networks - image processing, handwriting recognition
- many more
Further Reading
Backlinks:
MachineLearning:NeuralNetworks:BackPropagation:WeightInitializationAttachments:
ActivationFunction | 4.10kb | |
Appendix | 4.10kb | |
BackPropagation | 4.10kb | |
ConvolutionalNetwork | 4.10kb | |
diagram.dot | 536b | |
diagram.png | 12.4kb | |
diagram001.dot | 603b | |
diagram001.png | 18.1kb | |
diagram002.dot | 839b | |
diagram002.png | 21.5kb |