{
  "id": 383168,
  "title": "Is there a rule of thumb to choose an architecture for transfer learning?",
  "url": "/competitions/nfl-player-contact-detection/discussion/383168",
  "author_name": "Aishwary.Shukla_45",
  "post_date": "2023-02-02T13:45:39.349000",
  "votes": 0,
  "comment_count": 0,
  "views": 0,
  "content": "<p>I have one query. Today, there exists a large number of deep-learning models today in various frameworks. Each of which was developed for a specific purpose and performs better than its counterpart in certain scenarios and sometimes surprisingly in general scenarios as well. For a give problem, it happens at time that some architecture(s) may perfectly work while sometimes they don't. In this case faster experimentation becomes necessary and even there are times when performing a large set of experiments are not even possible. <br>\nGiven the above context, is there a way to roughly pre-determine that what architectures to start the trials with? For example, lets say for a classification problem with tabular datasets, algorithms like XGB, tree-based learning are some goto methods that most of them use and it works wonderfully well many times. So, are there any such architectures in deep learning as well?</p>\n<p>One thing to note here is that my query above was for image/video datasets give the competition. But, I am curious to know about other fields like text, audio, etc. as well.</p>\n<p>Any ideas/thoughts are appreciated.</p>\n<p>Also, if there is a need to improve the details in the query asked, please let me know. <br>\nThank you.</p>",
  "messages": [
    {
      "id": 2126841,
      "postDate": "2023-02-02T13:45:39.350Z",
      "content": "<p>I have one query. Today, there exists a large number of deep-learning models today in various frameworks. Each of which was developed for a specific purpose and performs better than its counterpart in certain scenarios and sometimes surprisingly in general scenarios as well. For a give problem, it happens at time that some architecture(s) may perfectly work while sometimes they don't. In this case faster experimentation becomes necessary and even there are times when performing a large set of experiments are not even possible. <br>\nGiven the above context, is there a way to roughly pre-determine that what architectures to start the trials with? For example, lets say for a classification problem with tabular datasets, algorithms like XGB, tree-based learning are some goto methods that most of them use and it works wonderfully well many times. So, are there any such architectures in deep learning as well?</p>\n<p>One thing to note here is that my query above was for image/video datasets give the competition. But, I am curious to know about other fields like text, audio, etc. as well.</p>\n<p>Any ideas/thoughts are appreciated.</p>\n<p>Also, if there is a need to improve the details in the query asked, please let me know. <br>\nThank you.</p>",
      "rawMarkdown": "I have one query. Today, there exists a large number of deep-learning models today in various frameworks. Each of which was developed for a specific purpose and performs better than its counterpart in certain scenarios and sometimes surprisingly in general scenarios as well. For a give problem, it happens at time that some architecture(s) may perfectly work while sometimes they don't. In this case faster experimentation becomes necessary and even there are times when performing a large set of experiments are not even possible. \nGiven the above context, is there a way to roughly pre-determine that what architectures to start the trials with? For example, lets say for a classification problem with tabular datasets, algorithms like XGB, tree-based learning are some goto methods that most of them use and it works wonderfully well many times. So, are there any such architectures in deep learning as well?\n\nOne thing to note here is that my query above was for image/video datasets give the competition. But, I am curious to know about other fields like text, audio, etc. as well.\n\nAny ideas/thoughts are appreciated.\n\nAlso, if there is a need to improve the details in the query asked, please let me know. \nThank you.\n"
    }
  ],
  "comments": [],
  "raw_markdown_by_id": {
    "2126841": "I have one query. Today, there exists a large number of deep-learning models today in various frameworks. Each of which was developed for a specific purpose and performs better than its counterpart in certain scenarios and sometimes surprisingly in general scenarios as well. For a give problem, it happens at time that some architecture(s) may perfectly work while sometimes they don't. In this case faster experimentation becomes necessary and even there are times when performing a large set of experiments are not even possible. \nGiven the above context, is there a way to roughly pre-determine that what architectures to start the trials with? For example, lets say for a classification problem with tabular datasets, algorithms like XGB, tree-based learning are some goto methods that most of them use and it works wonderfully well many times. So, are there any such architectures in deep learning as well?\n\nOne thing to note here is that my query above was for image/video datasets give the competition. But, I am curious to know about other fields like text, audio, etc. as well.\n\nAny ideas/thoughts are appreciated.\n\nAlso, if there is a need to improve the details in the query asked, please let me know. \nThank you.\n"
  }
}