Loading slide
Imagine pouring that 500 dollar vintage for Vincenza instead.
She might notice qualities the four-number model cannot represent. Perhaps it lasts longer on the palate or holds together in a way the teaching bottles did not. She may know she is drinking something unusual.
That still may not tell her what number to say.
Knowing a wine is unusual is not the same as knowing what unusual is worth. If none of her training bottles connected those qualities to a price, she lacks the relevant feedback.
She could hedge upward and still miss 500 dollars by a long way.
That is the same problem she had on her first day, arriving from the other direction. Back then the sensations meant nothing because she had no experience to attach to them. Here, one sensation still means nothing, for the same reason.
So the two failures are not quite the same, and the difference is worth holding onto.
The model cannot represent the missing quality at all. A person might notice it without knowing its price. Neither can reliably turn it into dollars without relevant experience, but only one may have a hint that the four recorded inputs are incomplete.
Which raises the obvious question, and it is the one that leads out of this chapter.
Could you just give the model more to work with?
Yes, and it is worth being exact about what that fixes. Add a unit for the year, another for the maker, another for the rarity, and the machine has seven units instead of four. Those three things now have wires, so they can reach the price. The model that could not see the vintage at all can now see it.
But seeing is not knowing. A fresh year unit arrives with a weight that means nothing, exactly like the four did at 8. And the eight bottles cannot teach it, because nobody ever recorded what year they were from. Teaching a seven-unit machine means starting again with bottles that carry all seven numbers, and enough variety among them, wines from ordinary years and famous ones, for the year weight to have anything to learn from.
# did you know?
Starting from scratch is not always necessary in practice. A model can often be extended and then trained further, keeping what its existing weights already know instead of throwing them away. That is a real technique, and it still needs examples carrying the new information.
So more units buy you the possibility of learning something. The learning still has to be paid for in examples, and that trade never goes away, however large the model gets.
Loading contents...