Deep Black BoxOpen up the black box of AI one problem at a time, with your own hands. 21 problems in all.
Question 5
The same function, now picking which one
That is the whole formula — get a score, turn it into a share
โ
Subtraction, so it falls to the right; addition, so it rises to the right; zero, so it is flat.
Move the hours of sleep and the answer switches between the three.
โ
Turned into shares that add up to 1 (softmax)
An LLM (an AI that writes text, like ChatGPT) is doing exactly the same thing.
Its job is guessing the next word, and the candidates number not three but 50,000.
The next word after “The capital of Japan is” →
Tokyo 82% / Osaka 11% / Kyoto 7% (100% in total)
These numbers are an example.
The 5.5 and 9.5 in the formula were placed by a person. But those two can also be
left to the machine, in the same way as Problem 2 (measure the effect, move the knob).
What to take away
This is classification. Put out one exit per candidate, and take the highest as the answer.
Inside, it is the same multiply and add as Problem 1; the only change is how the exits are attached.
Where this pays off
Whether you are predicting a number or picking one of several options changes the number of exits and what happens at the end.
next
What comes next
Next is Chapter 2, Deep learning. You add knobs and get your hands on what stacking layers makes possible.