Multi-Layer Perceptron (MLP) with single neuron implementation

Running instruction

• I used modelsim free edition that can be downloaded here
• Open project with MLP_Single.mpf

Simulation was conducted in ns.

Simulation

Simulation Setup

defined variables

• Number of Layer (M) is 3
• Number of Neuron each layer (N) is 3
• Fixed-point resolution 32 bit (12 integer part, 20 fractional part)
• Actication function is ReLu (will take only positive value, negative value will become 0)

Input data (X)

• X0 = 0.3
• X1 = 1.3
• X2 = 2.3

Weight data

Weight data is generated automatically for just simulate the calculation. W0 is used for weight bias, where the bias always 1. W1-Wn are used for weight input data.

Weight data for hidden layer

• W0_0 = 1
• W1_0 = 1.1
• W2_0 = 1.2
• W3_0 = 1.3
• W0_1 = 2
• W1_2 = 2.1
• W2_1 = 2.2
• W3_1 = 2.3
• W0_2 = 3
• W1_2 = 3.1
• W2_2 = 3.2
• W3_2 = 3.3

Weight data for output layer

• W0_0 = 11
• W1_0 = 11.1
• W2_0 = 11.2
• W3_0 = 11.3
• W0_1 = 12
• W1_2 = 12.1
• W2_1 = 12.2
• W3_1 = 12.3
• W0_2 = 13
• W1_2 = 13.1
• W2_2 = 13.2
• W3_2 = 13.3

Simulation result

Output from Hidden layer

• H0 = 5.88
• H1 = 10.78
• H2 = 15.68

Output from Output layer

• O0 = 374.188
• O1 = 407.528
• O2 = 440.868

Simulation from modelsim waveform

the simulation waveform can be loaded from file wave.do above.

The simulation result is confirmed by manual calculation from excel file in directory above.