{
  "id": 413976,
  "title": "Need help with embeddings.",
  "url": "/competitions/cafa-5-protein-function-prediction/discussion/413976",
  "author_name": "",
  "post_date": "2023-05-31T00:15:59.043317700Z",
  "votes": 4,
  "comment_count": 2,
  "views": 0,
  "content": "<p>Hi everyone,<br>\nI'm new to this compitition and I'm finding it difficult to understand the importance of protein embeddings.<br>\nSo how important are they? And should I create my own or just use the <code>t5embedds</code>?</p>",
  "messages": [
    {
      "id": "2281494",
      "postDate": "05/31/2023 00:15:59",
      "content": "<p>Hi everyone,<br>\nI'm new to this compitition and I'm finding it difficult to understand the importance of protein embeddings.<br>\nSo how important are they? And should I create my own or just use the <code>t5embedds</code>?</p>",
      "rawMarkdown": "Hi everyone,\nI'm new to this compitition and I'm finding it difficult to understand the importance of protein embeddings.\nSo how important are they? And should I create my own or just use the `t5embedds`?",
      "votes": null
    },
    {
      "id": "2281825",
      "postDate": "05/31/2023 06:40:18",
      "content": "<p>Imagine a 2D space, X and Y, and imagine one sentence like \"<strong>My book is new and cheap.</strong>\", so embeddings transform these words in vectors that you can put it in the 2D vector space, so in that case the vector new and cheap can be close together and \"my\" can be close to \"book\" or \"car\". </p>\n<p>Once you have created your embedding or vectors, you only have to \"<strong>trace lines</strong>\" in your vector space, and you can use one model for this purpose or even statistical methods, in order to fit or adjust what you need. </p>\n<p>The embeddings have two major problems, <strong>the first is that you are limited by the model that have created the embeddings and the second one you also are limited by vector overlaps.</strong></p>\n<p>In the first case, your model do not adjust the vectors how it should, and this is very important for the second part, if one word such as \"car\" is in the vector space of \"verbs\" for example the previous sentence \"My book car new and cheap\" does not have sense, and it is a problem for fitting, so, you always would be limited by the good performance of your model creating embeddings. </p>\n<p>And the second problem is that you are overlapping vectors, the second part \"fitting the model\" wouldn't be an easy task, the solution for this is creating bigger vector space, but memory can be a problem. If you separate more vectors each others, would be more easy for your model \"segment the meaning\" of your vectors, but in contrast vectors more bigger more memory would be needed.</p>",
      "rawMarkdown": "Imagine a 2D space, X and Y, and imagine one sentence like \"**My book is new and cheap.**\", so embeddings transform these words in vectors that you can put it in the 2D vector space, so in that case the vector new and cheap can be close together and \"my\" can be close to \"book\" or \"car\". \n\nOnce you have created your embedding or vectors, you only have to \"**trace lines**\" in your vector space, and you can use one model for this purpose or even statistical methods, in order to fit or adjust what you need. \n\nThe embeddings have two major problems, **the first is that you are limited by the model that have created the embeddings and the second one you also are limited by vector overlaps.**\n\nIn the first case, your model do not adjust the vectors how it should, and this is very important for the second part, if one word such as \"car\" is in the vector space of \"verbs\" for example the previous sentence \"My book car new and cheap\" does not have sense, and it is a problem for fitting, so, you always would be limited by the good performance of your model creating embeddings. \n\nAnd the second problem is that you are overlapping vectors, the second part \"fitting the model\" wouldn't be an easy task, the solution for this is creating bigger vector space, but memory can be a problem. If you separate more vectors each others, would be more easy for your model \"segment the meaning\" of your vectors, but in contrast vectors more bigger more memory would be needed.",
      "votes": null
    },
    {
      "id": "2284870",
      "postDate": "06/02/2023 10:14:24",
      "content": "<p>Ohh I see. Thanks a lot for the help.<br>\nSo I guess the key takeaway here would be:</p>\n<ol>\n<li>It's essential to have a powerful embedding model</li>\n<li>Enough memory and bigger vector space are good.</li>\n</ol>\n<p>Again, thanks!</p>",
      "rawMarkdown": "Ohh I see. Thanks a lot for the help.\nSo I guess the key takeaway here would be:\n1. It's essential to have a powerful embedding model\n2. Enough memory and bigger vector space are good.\n\nAgain, thanks!",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 2281825,
      "author_name": "victorfernandezalbor",
      "author_url": "",
      "post_date": "05/31/2023 06:40:18",
      "content": "<p>Imagine a 2D space, X and Y, and imagine one sentence like \"<strong>My book is new and cheap.</strong>\", so embeddings transform these words in vectors that you can put it in the 2D vector space, so in that case the vector new and cheap can be close together and \"my\" can be close to \"book\" or \"car\". </p>\n<p>Once you have created your embedding or vectors, you only have to \"<strong>trace lines</strong>\" in your vector space, and you can use one model for this purpose or even statistical methods, in order to fit or adjust what you need. </p>\n<p>The embeddings have two major problems, <strong>the first is that you are limited by the model that have created the embeddings and the second one you also are limited by vector overlaps.</strong></p>\n<p>In the first case, your model do not adjust the vectors how it should, and this is very important for the second part, if one word such as \"car\" is in the vector space of \"verbs\" for example the previous sentence \"My book car new and cheap\" does not have sense, and it is a problem for fitting, so, you always would be limited by the good performance of your model creating embeddings. </p>\n<p>And the second problem is that you are overlapping vectors, the second part \"fitting the model\" wouldn't be an easy task, the solution for this is creating bigger vector space, but memory can be a problem. If you separate more vectors each others, would be more easy for your model \"segment the meaning\" of your vectors, but in contrast vectors more bigger more memory would be needed.</p>",
      "votes": null,
      "replies": [
        {
          "id": 2284870,
          "author_name": "amitmawal",
          "author_url": "",
          "post_date": "06/02/2023 10:14:24",
          "content": "<p>Ohh I see. Thanks a lot for the help.<br>\nSo I guess the key takeaway here would be:</p>\n<ol>\n<li>It's essential to have a powerful embedding model</li>\n<li>Enough memory and bigger vector space are good.</li>\n</ol>\n<p>Again, thanks!</p>",
          "votes": null,
          "replies": []
        }
      ]
    }
  ],
  "raw_markdown_by_id": {
    "2281494": "Hi everyone,\nI'm new to this compitition and I'm finding it difficult to understand the importance of protein embeddings.\nSo how important are they? And should I create my own or just use the `t5embedds`?",
    "2281825": "Imagine a 2D space, X and Y, and imagine one sentence like \"**My book is new and cheap.**\", so embeddings transform these words in vectors that you can put it in the 2D vector space, so in that case the vector new and cheap can be close together and \"my\" can be close to \"book\" or \"car\". \n\nOnce you have created your embedding or vectors, you only have to \"**trace lines**\" in your vector space, and you can use one model for this purpose or even statistical methods, in order to fit or adjust what you need. \n\nThe embeddings have two major problems, **the first is that you are limited by the model that have created the embeddings and the second one you also are limited by vector overlaps.**\n\nIn the first case, your model do not adjust the vectors how it should, and this is very important for the second part, if one word such as \"car\" is in the vector space of \"verbs\" for example the previous sentence \"My book car new and cheap\" does not have sense, and it is a problem for fitting, so, you always would be limited by the good performance of your model creating embeddings. \n\nAnd the second problem is that you are overlapping vectors, the second part \"fitting the model\" wouldn't be an easy task, the solution for this is creating bigger vector space, but memory can be a problem. If you separate more vectors each others, would be more easy for your model \"segment the meaning\" of your vectors, but in contrast vectors more bigger more memory would be needed.",
    "2284870": "Ohh I see. Thanks a lot for the help.\nSo I guess the key takeaway here would be:\n1. It's essential to have a powerful embedding model\n2. Enough memory and bigger vector space are good.\n\nAgain, thanks!"
  },
  "source": "meta"
}