Are submissions that dont use any ML to perform this task allowed by the rules of the challenge?
\n\nI looked through the rules and FAQ and didn't see anything saying that using ML was a requirement. Perhaps I missed something.
\n\nI'd like to see if I can approach this problem with other analytical methods. Im new to Kaggle and this is my first challenge attempt, so sorry if this is a ridiculous question.
\n\nThanks!
", "messages": [ { "id": "714335", "postDate": "01/09/2020 10:09:28", "content": "Are submissions that dont use any ML to perform this task allowed by the rules of the challenge?
\n\nI looked through the rules and FAQ and didn't see anything saying that using ML was a requirement. Perhaps I missed something.
\n\nI'd like to see if I can approach this problem with other analytical methods. Im new to Kaggle and this is my first challenge attempt, so sorry if this is a ridiculous question.
\n\nThanks!
", "rawMarkdown": "Are submissions that dont use any ML to perform this task allowed by the rules of the challenge?\n\nI looked through the rules and FAQ and didn't see anything saying that using ML was a requirement. Perhaps I missed something.\n\nI'd like to see if I can approach this problem with other analytical methods. Im new to Kaggle and this is my first challenge attempt, so sorry if this is a ridiculous question.\n\nThanks!", "votes": null }, { "id": "714368", "postDate": "01/09/2020 11:07:37", "content": "What's your so-called no-ML-method? As far as I know, the leak has been removed.
", "rawMarkdown": "What's your so-called no-ML-method? As far as I know, the leak has been removed.", "votes": null }, { "id": "714376", "postDate": "01/09/2020 11:19:43", "content": "I'm not interested in the leak or any other form of taking advantage of unintended features of the data. I simply want to write an script that performs multiple different types of analysis on the video frames and audio to find some hallmark features of Deepfake modified video/audio.
", "rawMarkdown": "I'm not interested in the leak or any other form of taking advantage of unintended features of the data. I simply want to write an script that performs multiple different types of analysis on the video frames and audio to find some hallmark features of Deepfake modified video/audio.", "votes": null }, { "id": "714401", "postDate": "01/09/2020 12:01:56", "content": "Sounds interesting. I think you can use any type of methods.
", "rawMarkdown": "Sounds interesting. I think you can use any type of methods.", "votes": null }, { "id": "714728", "postDate": "01/09/2020 17:29:14", "content": "AFAIK the only requirement is that it must be automated. You can use non-ml computer vision algorithms for example
", "rawMarkdown": "AFAIK the only requirement is that it must be automated. You can use non-ml computer vision algorithms for example", "votes": null }, { "id": "714760", "postDate": "01/09/2020 18:11:40", "content": "The requirement is that you can make your submission through Kaggle's notebook/script, which supports Python, that it meets the code requirements, and that it is capable of being run on an unseen test set, as described in getting started. If you elect not to use traditional ML, that's fine, but you cannot (and will not benefit from) hand-label the test set.
", "rawMarkdown": "The requirement is that you can make your submission through Kaggle's notebook/script, which supports Python, that it meets the [code requirements](https://www.kaggle.com/c/deepfake-detection-challenge/overview/code-requirements), and that it is capable of being run on an unseen test set, as described in [getting started](https://www.kaggle.com/c/deepfake-detection-challenge/overview/getting-started). If you elect not to use traditional ML, that's fine, but you cannot (and will not benefit from) hand-label the test set.", "votes": null }, { "id": "714796", "postDate": "01/09/2020 19:02:12", "content": "Kevin - there are several of the shared kernels were folks made successful submissions while not actually doing anything from ML point of view. They like me (mine is not public) were test driving the submission process. They are all code that meets the code requirements - therefore - very much ok to the rules. They just mostly end up with bad log loss results :)
\n\nIf you can write a script that extracts usable information and can make better than random guesses you might have a OK log loss result.
\n\nBut for sure - if you took that same usable information and used it to train a ML method than you would end up with a better than OK result. For many of the competitions on Kaggle that I have been involved with there was almost always a none ML way to pick some of the really bad apples and perhaps some of the really good apples, but ML was needed to make all the other predictions in between.
\n\nIf you get a non-ML method that gets you into the Bronze metal stage - put an ad in the Team Wanted discussion post - folks will be jumping out of their computer chairs to get you on their team.
", "rawMarkdown": "Kevin - there are several of the shared kernels were folks made successful submissions while not actually doing anything from ML point of view. They like me (mine is not public) were test driving the submission process. They are all code that meets the code requirements - therefore - very much ok to the rules. They just mostly end up with bad log loss results :)\n\nIf you can write a script that extracts usable information and can make better than random guesses you might have a OK log loss result.\n\nBut for sure - if you took that same usable information and used it to train a ML method than you would end up with a better than OK result. For many of the competitions on Kaggle that I have been involved with there was almost always a none ML way to pick some of the really bad apples and perhaps some of the really good apples, but ML was needed to make all the other predictions in between. \n\nIf you get a non-ML method that gets you into the Bronze metal stage - put an ad in the Team Wanted discussion post - folks will be jumping out of their computer chairs to get you on their team.", "votes": null }, { "id": "714905", "postDate": "01/09/2020 22:59:29", "content": "Thanks so much for the answer and insights! And I agree that a ML model that uses the outputs of my analysis tools might give optimal results. That's what I intend to do, but I just wanted to make sure that if the model is not necessary, my submission still meets the requirements. Im super excited to begin testing and you just made my path forward much clearer. You rock!
", "rawMarkdown": "Thanks so much for the answer and insights! And I agree that a ML model that uses the outputs of my analysis tools might give optimal results. That's what I intend to do, but I just wanted to make sure that if the model is not necessary, my submission still meets the requirements. Im super excited to begin testing and you just made my path forward much clearer. You rock!", "votes": null }, { "id": "714922", "postDate": "01/09/2020 23:38:52", "content": "Fantastic. Thank you!
", "rawMarkdown": "Fantastic. Thank you!", "votes": null } ], "comments": [ { "id": 714368, "author_name": "feifeizaici", "author_url": "", "post_date": "01/09/2020 11:07:37", "content": "What's your so-called no-ML-method? As far as I know, the leak has been removed.
", "votes": null, "replies": [ { "id": 714376, "author_name": "kevinwillis", "author_url": "", "post_date": "01/09/2020 11:19:43", "content": "I'm not interested in the leak or any other form of taking advantage of unintended features of the data. I simply want to write an script that performs multiple different types of analysis on the video frames and audio to find some hallmark features of Deepfake modified video/audio.
", "votes": null, "replies": [] }, { "id": 714401, "author_name": "feifeizaici", "author_url": "", "post_date": "01/09/2020 12:01:56", "content": "Sounds interesting. I think you can use any type of methods.
", "votes": null, "replies": [] } ] }, { "id": 714728, "author_name": "arc144", "author_url": "", "post_date": "01/09/2020 17:29:14", "content": "AFAIK the only requirement is that it must be automated. You can use non-ml computer vision algorithms for example
", "votes": null, "replies": [] }, { "id": 714760, "author_name": "juliaelliott", "author_url": "", "post_date": "01/09/2020 18:11:40", "content": "The requirement is that you can make your submission through Kaggle's notebook/script, which supports Python, that it meets the code requirements, and that it is capable of being run on an unseen test set, as described in getting started. If you elect not to use traditional ML, that's fine, but you cannot (and will not benefit from) hand-label the test set.
", "votes": null, "replies": [ { "id": 714922, "author_name": "kevinwillis", "author_url": "", "post_date": "01/09/2020 23:38:52", "content": "Fantastic. Thank you!
", "votes": null, "replies": [] } ] }, { "id": 714796, "author_name": "pcjimmmy", "author_url": "", "post_date": "01/09/2020 19:02:12", "content": "Kevin - there are several of the shared kernels were folks made successful submissions while not actually doing anything from ML point of view. They like me (mine is not public) were test driving the submission process. They are all code that meets the code requirements - therefore - very much ok to the rules. They just mostly end up with bad log loss results :)
\n\nIf you can write a script that extracts usable information and can make better than random guesses you might have a OK log loss result.
\n\nBut for sure - if you took that same usable information and used it to train a ML method than you would end up with a better than OK result. For many of the competitions on Kaggle that I have been involved with there was almost always a none ML way to pick some of the really bad apples and perhaps some of the really good apples, but ML was needed to make all the other predictions in between.
\n\nIf you get a non-ML method that gets you into the Bronze metal stage - put an ad in the Team Wanted discussion post - folks will be jumping out of their computer chairs to get you on their team.
", "votes": null, "replies": [ { "id": 714905, "author_name": "kevinwillis", "author_url": "", "post_date": "01/09/2020 22:59:29", "content": "Thanks so much for the answer and insights! And I agree that a ML model that uses the outputs of my analysis tools might give optimal results. That's what I intend to do, but I just wanted to make sure that if the model is not necessary, my submission still meets the requirements. Im super excited to begin testing and you just made my path forward much clearer. You rock!
", "votes": null, "replies": [] } ] } ], "raw_markdown_by_id": { "714335": "Are submissions that dont use any ML to perform this task allowed by the rules of the challenge?\n\nI looked through the rules and FAQ and didn't see anything saying that using ML was a requirement. Perhaps I missed something.\n\nI'd like to see if I can approach this problem with other analytical methods. Im new to Kaggle and this is my first challenge attempt, so sorry if this is a ridiculous question.\n\nThanks!", "714368": "What's your so-called no-ML-method? As far as I know, the leak has been removed.", "714376": "I'm not interested in the leak or any other form of taking advantage of unintended features of the data. I simply want to write an script that performs multiple different types of analysis on the video frames and audio to find some hallmark features of Deepfake modified video/audio.", "714401": "Sounds interesting. I think you can use any type of methods.", "714728": "AFAIK the only requirement is that it must be automated. You can use non-ml computer vision algorithms for example", "714760": "The requirement is that you can make your submission through Kaggle's notebook/script, which supports Python, that it meets the [code requirements](https://www.kaggle.com/c/deepfake-detection-challenge/overview/code-requirements), and that it is capable of being run on an unseen test set, as described in [getting started](https://www.kaggle.com/c/deepfake-detection-challenge/overview/getting-started). If you elect not to use traditional ML, that's fine, but you cannot (and will not benefit from) hand-label the test set.", "714796": "Kevin - there are several of the shared kernels were folks made successful submissions while not actually doing anything from ML point of view. They like me (mine is not public) were test driving the submission process. They are all code that meets the code requirements - therefore - very much ok to the rules. They just mostly end up with bad log loss results :)\n\nIf you can write a script that extracts usable information and can make better than random guesses you might have a OK log loss result.\n\nBut for sure - if you took that same usable information and used it to train a ML method than you would end up with a better than OK result. For many of the competitions on Kaggle that I have been involved with there was almost always a none ML way to pick some of the really bad apples and perhaps some of the really good apples, but ML was needed to make all the other predictions in between. \n\nIf you get a non-ML method that gets you into the Bronze metal stage - put an ad in the Team Wanted discussion post - folks will be jumping out of their computer chairs to get you on their team.", "714905": "Thanks so much for the answer and insights! And I agree that a ML model that uses the outputs of my analysis tools might give optimal results. That's what I intend to do, but I just wanted to make sure that if the model is not necessary, my submission still meets the requirements. Im super excited to begin testing and you just made my path forward much clearer. You rock!", "714922": "Fantastic. Thank you!" }, "source": "meta" }