- #1
ADDA
- 67
- 2
I'm attempting to use a artificial neural network to store a fuzzy floating point variable. While writing the code, I became somewhat creative. I used the idea of an Set Reset Latch and statistically translated the hardware SR-latch into an artificial neural network. My mathematical goal is to determine if input is consecutively low or high. To achieve this, I've modified the neural network idea slightly. Since I did not have a way to calculate error (also a reason for thinking this up), I pointed the error vector down in terms of the difference of the second derivative and first derivative of the end layer's nodes, so that the network would converge on less and less change of nodes. Also I've modified this activation function:
1.0 / (1.0 + e^(-x))
to allow for three convergence areas: -1, 0, 1, as seen in the following output of kmPlot:
1.0 / (1.0 + e^(-x))
to allow for three convergence areas: -1, 0, 1, as seen in the following output of kmPlot: