-
## 一言でいうと
TransformerのAttentionについて、SpanをAdaptiveにしつつSparseにもする手法の提案。通常のsoftmaxでは割当確率がゼロにはならないので、これがノイズになる場合がある。そこで割当確率ゼロを許容するα-entmaxを用いている(=Sparseを可能にする)。ただ、softmaxとの性能差は微妙。
![image](https://u…
-
When Black Hole are present in the system, trajectories often show a somewhat non-differentiable behaviour reltated to a too small step-size in the numerical integration of the equations of motion. A …
-
## 一言でいうと
DNNとDecision Treeを複合させる研究。木の経路・分岐それぞれで表現学習を行う+木構造が固定的でないものは初とのこと。growthではleafにデータ分割/データ変換/既存維持のいずれかを追加・追加箇所以外の重みを固定し学習、を行なっていきlossが小さくなる操作を採用。refinementで全体の学習を行う
![image](https://user-…
-
Adapt StateParser to crossCompile using ScalaJs
labra updated
8 years ago
-
Adapt SWIPE to AVX2 and the 256-bit registers available in the new Intel Haswell CPUs to become available in June. Should allow 32-way SIMD parallelisation.
-
Hi,
Is it possible to adapt this to work with voipinnovations (the virtual number provider)? How difficult / easy would it be?
Sorry to post this on issues, found it the only way to reach you.
-
Once we've switched to make (#37) we can support adaptive workflows using recursive make. The prototypical use case is feature search, e.g. adding/removing one feature (or set of features) at a time a…
-
Hi @patrickrchao
Great work on the repo.
Just out of curiosity could this be adapted to test for distribution shift?
If yes how would this be done?
Cheers
Andrew
-
I'm not sure about the terminology here, but I sometimes find the ephemeris markers too dense AND too sparse at the same time when showing an ephemeris, because they are currently only showing a fixed…
-