Main characteristics: Only 1 layer (= input layer), no hidden layer. Feed-forward only. No Sigmoid activation function. No back propagation.

Learning is achieved simply by incrementally updating the connection weights based on the comparison with the desired target output (supervised learning).
Its performance (success rate) of 85% is far off the state-of-the-art techniques (surprise, surprise) but close the Yann Lecun's 88% when using only a linear classifier.

See also

Simple 1-Layer Neural Network for MNIST Handwriting Recognition

http://yann.lecun.com/exdb/mnist/

Version

Github Project Page

Author

Matt Lind

Date

July 2015