{
  "id": 36497,
  "title": "How to get Probability of Same Artist?",
  "url": "/competitions/painter-by-numbers/discussion/36497",
  "author_name": "",
  "post_date": "2017-07-17T12:25:35.210453300Z",
  "votes": null,
  "comment_count": 1,
  "views": 0,
  "content": "<p>I am tackling this problem by first training a convolutional neural network to learn what image was drawn by which artist. The output layer I expect would be some kind of probability mass function that outputs the probability that paintingi belongs to artist1, artist2, artist3, ....</p>\n\n<p>Once I have this information I can predict the probability that a painting was drawn by each artist. However, how can I convert the vectors of the probability mass function of painting1 and painting2 to get the probability that they are from the same artist?</p>\n\n<p>For example. lets say there are only 2 artists, the output of the convolutional neural network for painting1 returns [0.6, 0.4] (60% from artist1 and 40% from artist2) and the output for painting2 is [0.2, 0.8] (20% artist1 and 80% artist2). Using only this information how can I get a probability that they are the same?</p>\n\n<p>Am I on the right path or am I missing something very obvious.</p>",
  "messages": [
    {
      "id": "204026",
      "postDate": "07/17/2017 12:25:35",
      "content": "<p>I am tackling this problem by first training a convolutional neural network to learn what image was drawn by which artist. The output layer I expect would be some kind of probability mass function that outputs the probability that paintingi belongs to artist1, artist2, artist3, ....</p>\n\n<p>Once I have this information I can predict the probability that a painting was drawn by each artist. However, how can I convert the vectors of the probability mass function of painting1 and painting2 to get the probability that they are from the same artist?</p>\n\n<p>For example. lets say there are only 2 artists, the output of the convolutional neural network for painting1 returns [0.6, 0.4] (60% from artist1 and 40% from artist2) and the output for painting2 is [0.2, 0.8] (20% artist1 and 80% artist2). Using only this information how can I get a probability that they are the same?</p>\n\n<p>Am I on the right path or am I missing something very obvious.</p>",
      "rawMarkdown": "I am tackling this problem by first training a convolutional neural network to learn what image was drawn by which artist. The output layer I expect would be some kind of probability mass function that outputs the probability that paintingi belongs to artist1, artist2, artist3, ....\n\nOnce I have this information I can predict the probability that a painting was drawn by each artist. However, how can I convert the vectors of the probability mass function of painting1 and painting2 to get the probability that they are from the same artist?\n\nFor example. lets say there are only 2 artists, the output of the convolutional neural network for painting1 returns [0.6, 0.4] (60% from artist1 and 40% from artist2) and the output for painting2 is [0.2, 0.8] (20% artist1 and 80% artist2). Using only this information how can I get a probability that they are the same?\n\nAm I on the right path or am I missing something very obvious.",
      "votes": null
    },
    {
      "id": "204118",
      "postDate": "07/17/2017 16:30:02",
      "content": "<p>You might want to read about Siamese neural networks. They allow you to train an algorithm to examine two images and then predict whether they are by the same artist - without needing a step where the algorithm predicts which artist created each of the two paintings. Siamese neural networks also allow your algorithm to extrapolate to paintings by artists whose works the algorithm has not seen before.</p>\n\n<p><a href=\"https://github.com/small-yellow-duck/kaggle_art/blob/master/mnist_siamese_cnn.py\">https://github.com/small-yellow-duck/kaggle_art/blob/master/mnist_siamese_cnn.py</a></p>",
      "rawMarkdown": "You might want to read about Siamese neural networks. They allow you to train an algorithm to examine two images and then predict whether they are by the same artist - without needing a step where the algorithm predicts which artist created each of the two paintings. Siamese neural networks also allow your algorithm to extrapolate to paintings by artists whose works the algorithm has not seen before.\n\nhttps://github.com/small-yellow-duck/kaggle_art/blob/master/mnist_siamese_cnn.py",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 204118,
      "author_name": "smallyellowduck",
      "author_url": "",
      "post_date": "07/17/2017 16:30:02",
      "content": "<p>You might want to read about Siamese neural networks. They allow you to train an algorithm to examine two images and then predict whether they are by the same artist - without needing a step where the algorithm predicts which artist created each of the two paintings. Siamese neural networks also allow your algorithm to extrapolate to paintings by artists whose works the algorithm has not seen before.</p>\n\n<p><a href=\"https://github.com/small-yellow-duck/kaggle_art/blob/master/mnist_siamese_cnn.py\">https://github.com/small-yellow-duck/kaggle_art/blob/master/mnist_siamese_cnn.py</a></p>",
      "votes": null,
      "replies": []
    }
  ],
  "raw_markdown_by_id": {
    "204026": "I am tackling this problem by first training a convolutional neural network to learn what image was drawn by which artist. The output layer I expect would be some kind of probability mass function that outputs the probability that paintingi belongs to artist1, artist2, artist3, ....\n\nOnce I have this information I can predict the probability that a painting was drawn by each artist. However, how can I convert the vectors of the probability mass function of painting1 and painting2 to get the probability that they are from the same artist?\n\nFor example. lets say there are only 2 artists, the output of the convolutional neural network for painting1 returns [0.6, 0.4] (60% from artist1 and 40% from artist2) and the output for painting2 is [0.2, 0.8] (20% artist1 and 80% artist2). Using only this information how can I get a probability that they are the same?\n\nAm I on the right path or am I missing something very obvious.",
    "204118": "You might want to read about Siamese neural networks. They allow you to train an algorithm to examine two images and then predict whether they are by the same artist - without needing a step where the algorithm predicts which artist created each of the two paintings. Siamese neural networks also allow your algorithm to extrapolate to paintings by artists whose works the algorithm has not seen before.\n\nhttps://github.com/small-yellow-duck/kaggle_art/blob/master/mnist_siamese_cnn.py"
  },
  "source": "meta"
}