abstract_reasoning

l10n-placeholder1 は関係型のセット (プログレッション、XOR、OR、AND、一貫性のある結合) を示し、\\(O\\) はオブジェクトの型 (形状、線) を示し、\\(A\\) は属性の型 (サイズ、色、位置、数)。マトリックスの構造 \\(S\\)は、特定のマトリックスによってもたらされる課題を決定するトリプル \\(S={[r, o, a]}\\) のセットです。このデータセットを使用するには: ```python import tensorflow_datasets as tfds ds = tfds.load('abstract_reasoning', split='train') for ex in ds.take(4): print(ex) ``` [ガイドを参照してください。 [tensorflow_datasets](https://www.tensorflow.org/datasets) の詳細については、](https://www.tensorflow.org/datasets/overview) を参照してください。 " />
  • 説明:

手続き的に生成された行列 (PGM) のデータは、ニューラル ネットワークにおける抽象的推論の測定に関する論文、Barrett、Hill、Santoro らによるものです。 2018. 目標は、抽象的な推論に基づいてコンテキスト パネルから正しい答えを推測することです。

このデータ セットを使用するには、データ セット ページからすべての *.tar.gz ファイルをダウンロードし、~/tensorflow_datasets/abstract_reasoning/ に配置してください。

\(R\) は関係タイプのセット (プログレッション、XOR、OR、AND、一貫性のある結合) を表し、 \(O\) はオブジェクト タイプ (形状、線) を表し、 \(A\) は属性タイプ (サイズ、色、位置、番号)。マトリックス\(S\)の構造は、特定のマトリックスによってもたらされる課題を決定するトリプル \(S={[r, o, a]}\) のセットです。

スプリット
'test' 200,000
'train' 120万
'validation' 20,000
  • 機能構造:
FeaturesDict({
    'answers': Video(Image(shape=(160, 160, 1), dtype=uint8)),
    'context': Video(Image(shape=(160, 160, 1), dtype=uint8)),
    'filename': Text(shape=(), dtype=string),
    'meta_target': Tensor(shape=(12,), dtype=int64),
    'relation_structure_encoded': Tensor(shape=(4, 12), dtype=int64),
    'target': ClassLabel(shape=(), dtype=int64, num_classes=8),
})
  • 機能のドキュメント:
特徴クラスDtype説明
特徴辞書
答え動画(画像) (8、160、160、1) uint8
環境動画(画像) (8、160、160、1) uint8
ファイル名文章ストリング
メタターゲットテンソル(12) int64
relationship_structure_encodedテンソル(4,12) int64
目標クラスラベルint64
@InProceedings{pmlr-v80-barrett18a,
  title =    {Measuring abstract reasoning in neural networks},
  author =   {Barrett, David and Hill, Felix and Santoro, Adam and Morcos, Ari and Lillicrap, Timothy},
  booktitle =    {Proceedings of the 35th International Conference on Machine Learning},
  pages =    {511--520},
  year =     {2018},
  editor =   {Dy, Jennifer and Krause, Andreas},
  volume =   {80},
  series =   {Proceedings of Machine Learning Research},
  address =      {Stockholmsmassan, Stockholm Sweden},
  month =    {10--15 Jul},
  publisher =    {PMLR},
  pdf =      {http://proceedings.mlr.press/v80/barrett18a/barrett18a.pdf},
  url =      {http://proceedings.mlr.press/v80/barrett18a.html},
  abstract =     {Whether neural networks can learn abstract reasoning or whetherthey merely rely on superficial statistics is a topic of recent debate. Here, we propose a dataset and challenge designed to probe abstract reasoning, inspired by a well-known human IQ test. To succeed at this challenge, models must cope with various generalisation 'regimes' in which the training data and test questions differ in clearly-defined ways. We show that popular models such as ResNets perform poorly, even when the training and test sets differ only minimally, and we present a novel architecture, with structure designed to encourage reasoning, that does significantly better. When we vary the way in which the test questions and training data differ, we find that our model is notably proficient at certain forms of generalisation, but notably weak at others. We further show that the model's ability to generalise improves markedly if it is trained to predict symbolic explanations for its answers. Altogether, we introduce and explore ways to both measure and induce stronger abstract reasoning in neural networks. Our freely-available dataset should motivate further progress in this direction.}
}

abstract_reasoning/neutral (デフォルト設定)

  • 設定の説明: 両方の行列をエンコードする構造
    トレーニング セットとテスト セットには、 \(r \\in R\)-placeholder12 のトリプル \([r, o, a]\) が含まれています。
    \(o \\in O\)、および \(a \\in A\)。トレーニング セットとテスト セットはばらばらです。
    入力変数 (つまり、ピクセル) のレベルで発生する分離
    症状)。

  • データセットサイズ: 42.02 GiB

  • ( tfds.as_dataframe ):

abstract_reasoning/補間

  • 構成の説明: ニュートラル スプリットと同様に、 \(S\) は任意の
    トリプル \([r, o, a]\)。補間の場合、トレーニング セットで、
    属性が「色」または「サイズ」(つまり、順序付けられた属性)であった場合、値
    属性は、個別のセットの偶数インデックスのメンバーに制限されていました。
    一方、テスト セットでは、奇数インデックス値のみが許可されました。すべてのことに注意してください
    \(S\) には、 color または size 属性を持つトリプル \([r, o, a]\) placeholder18 が含まれていました。
    したがって、テスト セットのすべての質問に一般化が必要です。

  • データセットサイズ: 37.09 GiB

  • ( tfds.as_dataframe ):

abstract_reasoning/外挿

  • 構成の説明: 補間と同じですが、の値
    属性は、離散セットの下半分に制限されていました
    テストセットでは、上半分の値を取りました。

  • データセットサイズ: 35.91 GiB

  • ( tfds.as_dataframe ):

abstract_reasoning/attr.rel.pairs

  • 構成の説明: すべての \(S\) には、少なくとも 2 つのトリプルが含まれていました。
    \(([r_1,o_1,a_1],[r_2,o_2,a_2]) = (t_1, t_2)\)、そのうち 400 個が実行可能です。私達
    トレーニング セットに 360 個、テスト セットに 40 個をランダムに割り当てました。メンバー
    \((t_1, t_2)\) は、構造体\(S\) で一緒に発生しませんでした
    トレーニングセットにあり、すべての構造 \(S\) に少なくとも 1 つのそのようなペアがありました
    サブセットとしての\((t_1, t_2)\) 。

  • データセットサイズ: 41.07 GiB

  • ( tfds.as_dataframe ):

abstract_reasoning/attr.rels

  • 構成の説明: 私たちのデータセットには、29 の一意の可能性があります。
    トリプル \([r,o,a]\)。これらのうち 7 つをランダムにテスト セットに割り当てました。
    ただし、各属性はこのセットで 1 回だけ表されます。
    これらの差し出されたトリプルは、トレーニング セットの質問では決して発生しませんでした。
    テスト セット内のすべて \(S\) には、それらの少なくとも 1 つが含まれていました。

  • データセットサイズ: 41.45 GiB

  • ( tfds.as_dataframe ):

abstract_reasoning/attrs.pairs

  • 構成の説明: \(S\) には、少なくとも 2 つのトリプルが含まれていました。 20あります
    (順不同) 実行可能な属性のペア \((a_1, a_2)\)
    \(r_i, o_i, ([r_1,o_1,a_1],[r_2,o_2,a_2])\) は実行可能なトリプル ペアです
    \(([r_1,o_1,a_1],[r_2,o_2,a_2]) = (t_1, t_2)\).これらのペアのうち 16 個を割り当てました
    トレーニング用に 4 つ、テスト用に 4 つ。テスト セット内のペア \((a_1, a_2)\) の場合、
    トレーニング セットの\(S\) には、 \(a_1\) または \(a_2\)placeholder34 のトリプルが含まれていました。テストでは
    セット、すべての \(S\) には、 \(a_1\) および \(a_2\)placeholder37 のトリプルが含まれていました。

  • データセットサイズ: 40.98 GiB

  • ( tfds.as_dataframe ):

abstract_reasoning/attrs.shape.color

  • 構成の説明: 保留属性 shape-color。 \(S\)
    トレーニング セットには、 \(o\)=shape および \(a\)=color のトリプルは含まれていませんでした。
    テスト セット内のパズルを支配するすべての構造には、少なくとも 1 つのトリプルが含まれていました
    \(o\)=shape および \(a\)=color を使用します。

  • データセットサイズ: 41.21 GiB

  • ( tfds.as_dataframe ):

abstract_reasoning/attrs.line.type

  • 構成の説明: 保留属性の線種。 \(S\) で
    トレーニング セットには、 \(o\)=line および \(a\)=type のトリプルが含まれていませんでした。
    テスト セット内のパズルを支配するすべての構造には、少なくとも 1 つのトリプルが含まれていました
    \(o\)=line および \(a\)=type を使用します。

  • データセットサイズ: 41.40 GiB

  • ( tfds.as_dataframe ):