[HN Gopher] Efficient Transformer Knowledge Distillation: A Perf...
___________________________________________________________________
Efficient Transformer Knowledge Distillation: A Performance Review
Author : PaulHoule
Score : 52 points
Date : 2023-12-07 15:41 UTC (7 hours ago)
(HTM) web link (arxiv.org)
(TXT) w3m dump (arxiv.org)
| empath-nirvana wrote:
| I skimmed the paper but I don't really understand what knowledge
| generation actually entails.
| drawnwren wrote:
| https://arxiv.org/abs/1503.02531
| WhitneyLand wrote:
| Don't mean to be flip at all, may I suggest:
|
| 1. Use Gpt-4 with something like this:
|
| _"Help me understand what this paper is about and estimate
| whether the relevance and impact to the field is likely to be
| low, medium, or high.
|
| Explain jargon that may be specific to AI research, but don't
| bother explaining or expanding on terms familiar to a working
| software developer or basic undergraduate computer science."_
|
| 2. Follow the above prompt with either the abstract or the full
| text of the paper.
|
| 3. Post something useful here and save others the time.
|
| Do not - in my opinion - copy/paste LLM output as a comment.
| But would love to hear your own succinct, human sounding, HN
| guideline compatible thoughts.
| egnehots wrote:
| This paper combines knowledge distillation and efficient
| attention mechanisms.
|
| => It works (still efficient, lower cost).
|
| Not an unexpected result, but to their credit, they established a
| new benchmark to test these combinations. KD+LongFormer is one of
| the best ones, retaining 95.9% of the performance for 50.7% of
| the cost.
| mistrial9 wrote:
| Appendix A
|
| A.1 Data Collection Data for GONERD was obtained through Giant
| Oak's GONER software, which scraped web ar- ticles from public
| facing online news sources as well as the U.S. Department of
| Justice's justice.gov domain. This webtext data was randomly
| sampled with an upweighted probability toward documents from
| justice.gov so that justice.gov consisted of roughly 25% of the
| total GONERD dataset.
___________________________________________________________________
(page generated 2023-12-07 23:01 UTC)