Timeline for When training word2vec, why is a new negative sampling process formulated instead of simply downsampling?
Current License: CC BY-SA 4.0
9 events
| when toggle format | what | by | license | comment | |
|---|---|---|---|---|---|
| Oct 24 at 21:04 | comment | added | chicxulub | Relevant paper: Sigmoid Loss for Language Image Pre-Training arxiv.org/abs/2303.15343 | |
| Apr 25 at 21:08 | comment | added | chicxulub | It seems like there's a similar question for training other contrastive models, e.g., CLIP. See the discussion around categorical cross entropy vs (the sum of) binary cross entropy in github.com/openai/CLIP/issues/25 | |
| Sep 8, 2023 at 5:31 | history | edited | chicxulub | CC BY-SA 4.0 |
fix notation for re-expressing likelihood
|
| Sep 4, 2023 at 11:19 | history | edited | chicxulub | CC BY-SA 4.0 |
NS missing log
|
| Aug 14, 2023 at 23:23 | history | edited | chicxulub | CC BY-SA 4.0 |
writing
|
| Aug 14, 2023 at 13:12 | history | edited | chicxulub | CC BY-SA 4.0 |
clarify some things
|
| Aug 14, 2023 at 12:38 | history | edited | chicxulub | CC BY-SA 4.0 |
separate out sum so relationship b/t likelihoods is more obv
|
| Aug 13, 2023 at 22:18 | history | edited | chicxulub | CC BY-SA 4.0 |
clarify c
|
| Aug 13, 2023 at 21:48 | history | asked | chicxulub | CC BY-SA 4.0 |