מופשט_היגיון

l10n-placeholder1 מציין את קבוצת סוגי הקשר (התקדמות, XOR, OR, AND, איחוד עקבי), \\(O\\) מציין את סוגי האובייקטים (צורה, קו), ו-\\(A\\) מציין את סוגי התכונות (גודל, צבע, מיקום, מספר). המבנה של מטריצה, \\(S\\), הוא קבוצת השלשות \\(S={[r, o, a]}\\) הקובעים את האתגר שמציבה מטריצה ​​מסוימת. כדי להשתמש במערך נתונים זה: ```python ייבא tensorflow_datasets כ-tfds ds = tfds.load('abstract_reasoning', split='train') עבור ex ב-ds.take(4): print(ex) ``` ראה [המדריך ](https://www.tensorflow.org/datasets/overview) למידע נוסף על [tensorflow_datasets](https://www.tensorflow.org/datasets). " />
  • תיאור :

נתוני פרוצדורלי מטריצות (PGM) מתוך המאמר Measuring Abstract Reasoning in Neural Networks, Barrett, Hill, Santoro et al. 2018. המטרה היא להסיק את התשובה הנכונה מלוחות ההקשר בהתבסס על נימוק מופשט.

כדי להשתמש בערכת נתונים זו, אנא הורד את כל קבצי *.tar.gz מדף ערכת הנתונים והצב אותם ב-~/tensorflow_datasets/abstract_reasoning/.

\(R\) מציין את קבוצת סוגי הקשר (התקדמות, XOR, OR, AND, איחוד עקבי), \(O\) מציין את סוגי האובייקטים (צורה, קו), ו- \(A\) מציין את סוגי התכונות (גודל, צבע, מיקום, מספר). המבנה של מטריצה,\(S\), הוא קבוצה של משולשות \(S={[r, o, a]}\) שקובעות את האתגר שמציבה מטריצה ​​מסוימת.

לְפַצֵל דוגמאות
'test' 200,000
'train' 1,200,000
'validation' 20,000
  • מבנה תכונה :
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),
})
  • תיעוד תכונה :
תכונה מעמד צוּרָה Dtype תיאור
FeaturesDict
תשובות וידאו (תמונה) (8, 160, 160, 1) uint8
הֶקשֵׁר וידאו (תמונה) (8, 160, 160, 1) uint8
שם קובץ טֶקסט חוּט
meta_target מוֹתֵחַ (12,) int64
מבנה_יחס_מקודד מוֹתֵחַ (4, 12) int64
יַעַד ClassLabel 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 (תצורת ברירת המחדל)

  • תיאור תצורה : המבנים המקודדים את המטריצות בשני ה-
    ערכות הדרכה ובדיקה מכילות כל משולש \([r, o, a]\) עבור \(r \\in R\),
    \(o \\in O\), ו- \(a \\in A\). ערכות הדרכה ובדיקות מפורקות, עם
    הפרדה המתרחשת ברמת משתני הקלט (כלומר פיקסל
    ביטויים).

  • גודל מערך נתונים : 42.02 GiB

  • דוגמאות ( tfds.as_dataframe ):

מופשט_היגיון/אינטרפולציה

  • תיאור תצורה : כמו בפיצול הנייטרלי, \(S\) מורכב מכל
    משולש \([r, o, a]\). לאינטרפולציה, בסט האימונים, כאשר ה
    התכונה הייתה "צבע" או "גודל" (כלומר, התכונות המסודרות), הערכים של
    התכונות הוגבלו לחברים בעלי אינדקס זוגי של קבוצה בדיד,
    ואילו במערך הבדיקות הותרו רק ערכים בעלי אינדקס אי זוגי. שימו לב שהכל
    \(S\) הכיל כמה \([r, o, a]\) משולש עם התכונה color או size .
    לפיכך, נדרשת הכללה לכל שאלה במערך המבחן.

  • גודל מערך נתונים : 37.09 GiB

  • דוגמאות ( tfds.as_dataframe ):

מופשט_היגיון/אקסטרפולציה

  • תיאור תצורה : זהה לאינטרפולציה, אבל הערכים של
    התכונות הוגבלו לחצי התחתון של הסט הדיסקרטי במהלך
    אימון, ואילו במערך המבחן הם לקחו ערכים בחצי העליון.

  • גודל מערך נתונים : 35.91 GiB

  • דוגמאות ( tfds.as_dataframe ):

abstract_reasoning/attr.rel.pairs

  • תיאור תצורה : כל \(S\) הכילו לפחות שני שלשות,
    \(([r_1,o_1,a_1],[r_2,o_2,a_2]) = (t_1, t_2)\), מתוכם 400 ברי קיימא. אָנוּ
    הוקצו באופן אקראי 360 למערך האימונים ו-40 למערך המבחנים. חברים
    \((t_1, t_2)\) מתוך 40 הזוגות המוחזקים לא התרחשו יחד במבנים\(S\)
    במערך האימונים, ולכל המבנים \(S\) היה לפחות זוג אחד כזה
    \((t_1, t_2)\) כקבוצת משנה.

  • גודל מערך נתונים : 41.07 GiB

  • דוגמאות ( tfds.as_dataframe ):

abstract_reasoning/attr.rels

  • תיאור תצורה : במערך הנתונים שלנו, ישנם 29 ייחודיים אפשריים
    משולש \([r,o,a]\). הקצהנו שבעה כאלה למערך המבחנים, באקראי,
    אבל כך שכל אחת מהתכונות הייתה מיוצגת פעם אחת בדיוק בקבוצה הזו.
    השלשות המתוחזקות הללו מעולם לא התרחשו בשאלות במערך האימונים, ו
    כל \(S\) בערכת הבדיקה הכיל לפחות אחד מהם.

  • גודל ערכת נתונים : 41.45 GiB

  • דוגמאות ( tfds.as_dataframe ):

abstract_reasoning/attrs.pairs

  • תיאור תצורה : \(S\) הכיל לפחות שני שלשות. יש 20
    זוגות קיימא (לא מסודרים) של תכונות \((a_1, a_2)\) כך שעבור חלק
    \(r_i, o_i, ([r_1,o_1,a_1],[r_2,o_2,a_2])\) הוא זוג משולש בר-קיימא
    \(([r_1,o_1,a_1],[r_2,o_2,a_2]) = (t_1, t_2)\). הקצנו 16 מהזוגות האלה
    לאימון וארבעה לבדיקה. עבור זוג \((a_1, a_2)\) בערכת הבדיקה,
    \(S\) בערכת ההדרכה הכיל שלשות עם \(a_1\) או \(a_2\). במבחן
    קבוצה, כל \(S\) הכילו שלשות עם \(a_1\) ו- \(a_2\).

  • גודל מערך נתונים : 40.98 GiB

  • דוגמאות ( tfds.as_dataframe ):

abstract_reasoning/attrs.shape.color

  • תיאור תצורה : תכונת צורה-צבע מוחזקת. \(S\) אינץ'
    ערכת ההדרכה לא הכילה שלשות עם \(o\)=צורה ו- \(a\)=צבע.
    כל המבנים השולטים בחידות בערכת המבחן הכילו לפחות משולש אחד
    עם \(o\)=צורה ו- \(a\)=צבע.

  • גודל מערך נתונים : 41.21 GiB

  • דוגמאות ( tfds.as_dataframe ):

abstract_reasoning/attrs.line.type

  • תיאור תצורה : סוג קו של תכונה מוחזק. \(S\) אינץ'
    ערכת ההדרכה לא הכילה שלשות עם \(o\)=line ו- \(a\)=type.
    כל המבנים השולטים בחידות בערכת המבחן הכילו לפחות משולש אחד
    עם \(o\)=קו ו- \(a\)=type.

  • גודל מערך נתונים : 41.40 GiB

  • דוגמאות ( tfds.as_dataframe ):