This is about a neural network utilized to test dosages of a medicine.
It's a demo based on a YouTube video, "Neural Networks Pt. 1: Inside the Black Box" by StatQuest with Josh Starmer.
The author does a great job explaining math details of a simple forward propagation using softplus activation.
This project is to enable discussion within that video community to resolve a problem mentioned below.
Apparently in a previous operation, a training dataset with labels has been utilized to obtain parameters that classify dosages of a medication.
The premise is that low and high dosages result in efficacy of 0 (medicine does not work), and medium dosages result in efficacy of 1 (medicine does work).
I think in most NNs, the testing phase does not utilize the steps of propagation.
But to explain the math of forward propagation, the video makes one pass to produce a result.
In my demo, I have dosages of (0.1, 0.3, 0.5, 0.7, and 0.9).
The network has two nodes in one middle layer.
The video shows two blue curved lines for the top node, and two orange curved lines for the bottom node.
I call these g1 and g2. The activation function produces g2 values.
The two g2 values are summed to g3 values, the green curve line.
A final negative bias is added to produce the final efficacy value.
Before the python script, I worked out the math in Excel.
I was able to confirm math results with a few given values in the video.
The python script delivers equal results for those few values.
It would be helpful if the video provided more math results, say from the activation (I call g1).
The script prints values for top_g1_y, top_g2_y, g3_y, and eff_y.
A Matplotlib plot is generated with six plots:
- top g1 and g2 (blue)
- bottom g1 and g2 (orange)
- sum g3 (green)
- efficacy (red)
The green and red lines are supposed to be squiggly lines, centered on y values of 0 and 1. The medium dosage value (0.5) is the only one that works as planned, with efficacy of 1. My code does not produce squiggly lines as shown in the video, to show both the low and high dosage values close to zero.