private void GenerateNetwork()
{
//create neurons
var output = new FiringNeuron[Options.LayerStructure.Length, Options.MaxLayerDensity];
for (var layer = 1; layer < Options.LayerStructure.Length; layer++)
{
for (var neuron = 0; neuron < Options.LayerStructure[layer]; neuron++)
{
output[layer, neuron] =
new FiringNeuron(new Coordinate(layer, neuron), Options.LayerStructure[layer - 1], new SigmoidActivationFunction());
}
}
NeuronLayers = output;
_proposedNeurons = new IProposedNeuron[Options.LayerStructure.Length, Options.MaxLayerDensity];
}
public double GetActivationSlopeAt(FiringNeuron neuron)
{
var tanh = neuron.Output;
return(1 - tanh * tanh);
}
public override double GetActivationSlopeAt(FiringNeuron neuron) => neuron.Output * (1 - neuron.Output);
public override double GetActivationSlopeAt(FiringNeuron neuron)
{
double y = neuron.Output;
return(y * (1 - y));
}
// This is the derivative after modifying the loss function. public override double GetActivationSlopeAt(FiringNeuron neuron) => 1;
public override double GetActivationSlopeAt(FiringNeuron neuron) => neuron.TotalInput > 0 ? 1 : .01;
public abstract double GetActivationSlopeAt(FiringNeuron neuron);
public static void Main()
{
int totalSamples = 1000;
double learningRate = 0.02;
#region Manufacture some data
var random = new Random();
var data = (
from i in Enumerable.Range(0, totalSamples)
let input1 = random.Next(2)
let input2 = random.Next(2)
select new
{
input1,
input2,
// A NAND gate should be 0 only when both inputs are 1
DesiredOutput = input1 == 1 && input2 == 1 ? 0 : 1
}).ToArray();
#endregion
// Split the data into training and testing sets
int trainingCount = totalSamples * 8 / 10;
var trainingSet = data.Take(trainingCount);
var testingSet = data.Skip(trainingCount);
var neuron = new Neuron();
var firingNeuron = new FiringNeuron(neuron);
// Train
foreach (var sample in trainingSet)
{
firingNeuron.Learn(sample.input1, sample.input2, sample.DesiredOutput, learningRate);
}
// Test
var testResults = (
from sample in testingSet
let actualOutput = firingNeuron.Fire(sample.input1, sample.input2)
let success = actualOutput >= .5 == (sample.DesiredOutput == 1)
let error = actualOutput - sample.DesiredOutput
group error by success).ToArray();
// Report results
neuron.PrintDump();
testResults
.Select(t => new { Successful = t.Key, Count = t.Count() })
.PrintDump();
testResults
.SelectMany(t => t)
.Average(t => Math.Abs(t))
.PrintDump();
testResults
.SelectMany(t => t)
.GroupBy(Loss => Math.Round(Loss, 2))
.Select(g => new { Error = g.Key, Count = g.Count() })
.OrderBy(g => g.Error)
.PrintDump();
}
