应用神经网络识别数字

Question

我试着去理解神经网络

我将输入数组合为

..# ### ### #.#
.## ..# ..# #.#
..# ### ### ###
..# #.. ..# ..#
..# ### ### ..#, etc

期望的ouptut我设置为数字/ 10,即数字= 5输出= 0.5

代码

require 'ruby-fann'

train = RubyFann::TrainData.new(
  inputs: [
    [0,0,1,0,1,1,0,0,1,0,0,1,0,0,1],
    [1,1,1,0,0,1,1,1,1,1,0,0,1,1,1],
    [1,1,1,0,0,1,1,1,1,0,0,1,1,1,1],
    [1,0,1,1,0,1,1,1,1,0,0,1,0,0,1],
    [1,1,1,1,0,0,1,1,1,0,0,1,1,1,1],
    [1,1,1,1,0,0,1,1,1,1,0,1,1,1,1],
    [1,1,1,0,0,1,0,1,0,1,0,0,1,0,0],
    [1,1,1,1,0,1,1,1,1,1,0,1,1,1,1],
    [1,1,1,1,0,1,1,1,1,0,0,1,1,1,1]
  ],
  desired_outputs: [[0.1],[0.2],[0.3], [0.4], [0.5], [0.6], [0.7], [0.8], [0.9]]
)
fann = RubyFann::Standard.new(
  num_inputs: 15,
  hidden_neurons: [8,4,3,4,1],
  num_outputs: 1
)
fann.train_on_data(train, 100000, 10, 0.1) # 100000 max_epochs, 100 errors between reports and 0.1 desired MSE (mean-squared-error)
outputs = fann.run([0,0,1,0,1,1,0,0,1,0,0,1,0,0,1])
result = outputs.first
abort result.inspect

每个运行脚本的输出

0.5367386954219215
0.5141728468011051
0.5249739971144654
0.5373135467504666
0.5182686028674102
0.46710004502372293
0.4723526462690119
0.5306690734137796
0.5151398228322749
0.5359153267266001
0.469100790593523
0.4749347798092478
0.5094355973839471
0.5205985468860461
0.5277528652471375
0.4825827561254995

我不明白为什么输出不等于0.1,这与第一次输入完全相同.

什么意味着价值在0.46 - 0.53 diapason？

UPDATE

我用0.1替换0,用0.9替换1

产量

0.4794515462681635
0.5332274595769928
0.4601992972516728
0.427064909364266
0.43466252163025687
0.46931411920827737
0.4455544021835517
0.48051179013023565
0.4798245565677274
0.4479353078492235
0.4646710791032779
0.4887400910135108

此外,我为零位添加+1输入,没有发生任何重大事件

Answer 1

训练神经网络是一种黑暗的艺术.在这里,您最大的问题是将RMS误差目标设置为0.1 - 这意味着您将接受大于您感兴趣的差异的平均绝对误差.将其设置得更低应该会有很大帮助.

另外(但不太重要):

• 你不需要那么多隐藏层.足够的神经元.从反复试验中我认为你的[8,4,3,4,1]问题有点低(最后一个没有用).价值[30]似乎有效 - 我基本上通过尝试一些猜测得到了这个.

• 分类通常最好使用每个类别一个0/1输出,然后选择最大值.你不需要那个,我测试你的0.1,0.2等目标,它就像那样工作得很好.解释为什么单独的输出更好:如果你的输入有一些噪音,理想情况下网络应该在3和8之间选择,那么使用单个输出的中间值可能是0.55 - 即使你绕它也不是很有用价值基本上是不正确的.但是,有9个输出用于分类,"3"和"8"的输出都会很高,你可以选择略高的输出,或者有信心地显示正确的值是"3" "或"8".

• 您选择作为测试用例的问题可能会陷入局部最小值,您需要调整动量和学习率以获得更好的成功机会.

• 改组训练数据也可能有所帮助.

您的代码/参数的以下更改应该更接近您期望的结果:

require 'ruby-fann'

train = RubyFann::TrainData.new(
  inputs: [
    [0,0,1,0,1,1,0,0,1,0,0,1,0,0,1], [1,1,1,0,0,1,1,1,1,1,0,0,1,1,1],
    [1,1,1,0,0,1,1,1,1,0,0,1,1,1,1], [1,0,1,1,0,1,1,1,1,0,0,1,0,0,1],
    [1,1,1,1,0,0,1,1,1,0,0,1,1,1,1], [1,1,1,1,0,0,1,1,1,1,0,1,1,1,1],
    [1,1,1,0,0,1,0,1,0,1,0,0,1,0,0], [1,1,1,1,0,1,1,1,1,1,0,1,1,1,1],
    [1,1,1,1,0,1,1,1,1,0,0,1,1,1,1]
  ],
  desired_outputs: [ [1,0,0,0,0,0,0,0,0], [0,1,0,0,0,0,0,0,0],
                     [0,0,1,0,0,0,0,0,0], [0,0,0,1,0,0,0,0,0],
                     [0,0,0,0,1,0,0,0,0], [0,0,0,0,0,1,0,0,0],
                     [0,0,0,0,0,0,1,0,0], [0,0,0,0,0,0,0,1,0],
                     [0,0,0,0,0,0,0,0,1] ]
)

fann = RubyFann::Standard.new(
  num_inputs: 15,
  hidden_neurons: [30],
  num_outputs: 9
)

fann.learning_rate = 0.5
fann.momentum = 0.5

fann.train_on_data(train, 10000, 1000, 0.001)

outputs = fann.run([0,0,1,0,1,1,0,0,1,0,0,1,0,0,1])
m = outputs.max
puts "Result: #{( outputs.find_index { |x| x == m } ) + 1}"