Closed yifliu3 closed 8 months ago
Hi, apologies for the delayed response.
In short, we optimize a discrete language feature space $\mathcal{S}$ using the loss (7) as described in the second and last paragraph of subsection 3.3.
Based on my understanding,the discrete feature set is optimized the loss defined in Eq.5. I am wondering if my understandings are right or not, hope you can solve my problems. Thanks a lot.
Hi, we optimize the discrete language feature space $\mathcal{S}$ using loss (7), which is formulated as a combination of losses (5) and (6).
For an in-depth understanding of the impact and specifics of loss (6), please refer to the ‘Optimization’ section under subsection 3.3, as well as section 7 in the supplementary material.
Dear authors,
It seems that the quantization procedure is optimized by the defined loss, but I'm confused about the optimizing object. Is it a MLP, or others? I cannot find in the paper.