-
## 一言でいうと
単語埋め込み行列の圧縮について。例えばdogとdogsで別々にとっているのはばからしいので、埋め込み行列を一つではなく複数の行列(=コードブック)の組み合わせで表現することを提案。さらに、これをAutoEncoderのような形式で既存の埋め込み行列から自動的に学習する。元の精度を維持し94%以上の圧縮を達成。
![image](https://user-images…
-
## 一言でいうと
複数のベース戦略を組み合わせ、様々な戦略を実現する手法の提案。通常、戦略の組み合わせは合算(Additive)で行われることが多いが、掛け合わせ(Multiplicative)を使うことで複数戦略を統合して行動分布を作るような形をとっている。これにより、複雑な連続コントロールタスクができることを確認
### 論文リンク
https://arxiv.org/ab…
-
## 一言でいうと
DNNの実行で、再利用可能なモジュールを組み合わせる試み。全体としては強化学習(AlphaZero)の枠組みで、環境以外にプログラムのembeddingを受け取りLSTMで処理、Action/Valueを出力するという流れになっている。LSTMはプログラムの実行環境として動作し、STOPというactionが出るまで実行を続ける。
![image](https://u…
-
## 一言でいうと
模倣学習においてエキスパートの行動を潜在表現にする際、行動をパートに分解するという手法(物をつかむ/持ち上げる/移動する、など)。基本はVAEで、Encodeを行う際にEncodeパートの境界も予測させる(境界数は指定する)。学習させる際は、予測対象外(境界外)にMaskをかけてBackpropを行う。
![image](https://user-images.gi…
-
can't find Siamese Contrastive Embedding Network for Compositional Zero-Shot Learning CVPR2022
-
Algorithm 2: Particle learning from Grosse, R. _Model selection in compositional spaces_. Ph.D. thesis, 2014.
fsaad updated
8 years ago
-
Hi,
Thanks for the impressive idea and high-quality code.
I am trying to experiment on the compositional learning with CelebA dataset. However, I just found the path for CelebA is defined in [t…
-
Sometimes we know that a set of variables should add up to a given total. Measurements involving proportions, percentages, probabilities, concentrations are compositional data. These data occur often …
-
Have you tought to target a curate StableHLO operators set?
https://github.com/openxla/stablehlo
Historically one of the main issue about Deep learning frameworks was the over the tome combinatorial …
bhack updated
11 months ago
-
Description:
The goal of this issue is to create a foundation for developing a cybernetic organism or open game using physics-informed syntax. To begin, we need to implement a basic active inference …