[HN Gopher] Controlling GPT-3 with Logit Bias
       ___________________________________________________________________
        
       Controlling GPT-3 with Logit Bias
        
       Author : ve55
       Score  : 40 points
       Date   : 2021-01-25 14:19 UTC (8 hours ago)
        
 (HTM) web link (medium.com)
 (TXT) w3m dump (medium.com)
        
       | sweezyjeezy wrote:
       | Not sure if useful - but if you know the initial biases of
       | outputs, you can recalibrate these yourselves, provided you have
       | output probabilities for all tokens (or all non-negligible ones
       | at least).
       | 
       | Say the model outputs n tokens, and the prior (bias) in the model
       | for tokens is m = (m_1, m_2, .... m_n), and the new prior you
       | want is n=(n_1, n_2, ..._)
       | 
       | Then if the model outputs prediction p = (p_1, ... , p_n) for all
       | tokens, then the new output you are looking for is
       | 
       | bias_shift(p) = softmax(logit(p) + log(n) - log(m))
       | 
       | You can prove this using Bayes rule + a bit of algebra. Most ML
       | people don't seem to know this trick, but it's super useful for
       | domain adaptation / class upsampling, where you know that the
       | class balance in your training set is different to the one you
       | want to predict on.
        
         | albertzeyer wrote:
         | This is like the density ratio approach
         | (https://arxiv.org/abs/2002.11268). The Hybrid Autoregressive
         | Transducer also uses a similar idea
         | (https://arxiv.org/abs/2003.07705).
         | 
         | Do you really need the renormalization, i.e. the outer softmax?
        
         | gradys wrote:
         | This is a great tip, thanks! Is this essentially importance
         | sampling? Reminds me of a technique for off-policy RL.
        
           | 317070 wrote:
           | They are related, but not the same. This recalibrating of
           | logits is basically what you would do afterwards, to correct
           | for the fact that you have been importance sampling your
           | replay buffer.
        
       ___________________________________________________________________
       (page generated 2021-01-25 23:02 UTC)