3 ms·

How is the selected token "random" if it's chosen based on weighted probabilities, especially if you use a temperature of 0?

by Timon3 3y ago

Karellen 3y ago
Because even with weighted probabilities, they're still probabilities. If you have "the cat sat on the", and the next possible tokens are "mat" with a probability of 90%, "floor" with a probability of 9%, and "dog" at 1%, how are you going to pick "mat" 90% of the time without consulting a (P)RNG?
- Timon3 3y ago
  They are only probabilities if you treat them as such. If you always just pick the highest probability option (temperature = 0), it's not random in any way - then it's just a confidence value. You can use randomness, but it's in no way inherent to the process.
  - Karellen 3y ago
    > If you always just pick the highest probability option (temperature = 0), it's not random in any way Huh. That feels like a recipe for either outputting parts of the training data verbatim, or getting stuck in an output loop. Although I earlier gave an example with 90/9/1 weights where the obvious choice is significantly higher than the others, I'd have thought that in a different scenario where you have, say, 51/49 or 34/33/32 options to choose from, always picking the one with only a fractionally higher confidence than the reasonable alternatives could lead to a "sameyness" in the outputs. Isn't there any value in having variance in the generated responses?
    - Timon3 3y ago
      > Isn't there any value in having variance in the generated responses? Sure, which is why the technique is used. But that doesn't mean "random" is a useful descriptor. I, as a human, try to add variance to my responses as well - does that mean my mind is "random"? That's why my point is that there is no inherent randomness in the process. We use a bit of randomness to make it subjectively better, but it's fundamentally not random.

3 ms·

How is the selected token "random" if it's chosen based on weighted probabilities, especially if you use a temperature of 0?

by Timon3 3y ago

How is the selected token "random" if it's chosen based on weighted probabilities, especially if you use a temperature of 0?

Karellen 3y ago
Because even with weighted probabilities, they're still probabilities. If you have "the cat sat on the", and the next possible tokens are "mat" with a probability of 90%, "floor" with a probability of 9%, and "dog" at 1%, how are you going to pick "mat" 90% of the time without consulting a (P)RNG?
- Timon3 3y ago
  They are only probabilities if you treat them as such. If you always just pick the highest probability option (temperature = 0), it's not random in any way - then it's just a confidence value. You can use randomness, but it's in no way inherent to the process.
  - Karellen 3y ago
    > If you always just pick the highest probability option (temperature = 0), it's not random in any way Huh. That feels like a recipe for either outputting parts of the training data verbatim, or getting stuck in an output loop. Although I earlier gave an example with 90/9/1 weights where the obvious choice is significantly higher than the others, I'd have thought that in a different scenario where you have, say, 51/49 or 34/33/32 options to choose from, always picking the one with only a fractionally higher confidence than the reasonable alternatives could lead to a "sameyness" in the outputs. Isn't there any value in having variance in the generated responses?
    - Timon3 3y ago
      > Isn't there any value in having variance in the generated responses? Sure, which is why the technique is used. But that doesn't mean "random" is a useful descriptor. I, as a human, try to add variance to my responses as well - does that mean my mind is "random"? That's why my point is that there is no inherent randomness in the process. We use a bit of randomness to make it subjectively better, but it's fundamentally not random.