Locally Connected layer

Presentation

This layer is mainly used to reduce the size of input, the Locally Connected layer makes it easier to extract markers from the images. The Locally Connected layer can be 1D or 2D depending on the previous layer.

Declaration

This is the function used to declare a Locally Connected layer.

template <class ... TOptimizer>
LocallyConnected(int numberOfLocallyConnected, int sizeOfLocalMatrix, activation activation = activation::sigmoid, TOptimizer ... optimizers);

Arguments

• numberOfLocallyConnected: The number of output filters. Multiply the number of neurons.
• sizeOfLocalMatrix: The size of the matrix. For a 2D convolution the matrix is a square of length sizeOfLocalMatrix.
• activation: The activation function of the neurons of the layer. See list of activation function

Here is an example of neural networks with 2D input. The Locally Connected layer receives a shape of 28 x 28 x 1 input and outputs a size of 7 x 7 x 2 output.

 StraightforwardNeuralNetwork neuralNetwork({
        Input(28, 28, 1),
        LocallyConnected(2, 4),
        FullyConnected(150),
        FullyConnected(70),
        FullyConnected(10)
    });

See an example of Locally Connected layer on dataset