Hi guys, I was just wondering, maybe someone of you could explain it to me:
\n\nWhat happens if we first train a multi instance learning model with 12 tiles. Later we decide to increase the number of tiles to 24. The backbone of the model should already have learned some features in my opinion from always using 12 tiles. Or will it be complete garbage and just rewrite the weights?
\n\nObviously, the main difference will be in the classifier, after the backbone.
", "messages": [ { "id": 875052, "postDate": "2020-06-05T13:45:23.417Z", "content": "This should be similar to doing progressive size learning where you start with a small image to kickstart your model and then use bigger images. It can lead to better performance for image classification so it might be worth looking into here also but by playing with the num_tiles instead of image size. Maybe more white areas can create issues though.
\n\nOn the other hand the major difference is that here you'd be first not giving all the information which could lead to higher overfit if the model learns irrelevant features when the cancer tiles are not included.
", "rawMarkdown": "This should be similar to doing progressive size learning where you start with a small image to kickstart your model and then use bigger images. It can lead to better performance for image classification so it might be worth looking into here also but by playing with the num_tiles instead of image size. Maybe more white areas can create issues though.\n\nOn the other hand the major difference is that here you'd be first not giving all the information which could lead to higher overfit if the model learns irrelevant features when the cancer tiles are not included.", "votes": 1, "replies": [ { "id": 875517, "postDate": "2020-06-05T21:43:17.800Z", "content": "Excellent answer, thank you very much!
\n\nThese are valid points. I also believe that a critical problem is that information was withheld and thus could have learned the wrong things.
", "rawMarkdown": "Excellent answer, thank you very much!\n\nThese are valid points. I also believe that a critical problem is that information was withheld and thus could have learned the wrong things." }, { "id": 875556, "postDate": "2020-06-05T22:55:29.917Z", "content": "Is it possible to train a model with a variable number of tiles between the examples?
", "rawMarkdown": "Is it possible to train a model with a variable number of tiles between the examples?" }, { "id": 876654, "postDate": "2020-06-06T21:46:38.657Z", "content": "@yannmajewski Only if you have a pooling function that can deal with the varrying length. Max pools, attention, LSTM, etc.
\n\nAverage pooling though cannot deal with it (in theory).
", "rawMarkdown": "@yannmajewski Only if you have a pooling function that can deal with the varrying length. Max pools, attention, LSTM, etc.\n\nAverage pooling though cannot deal with it (in theory)." }, { "id": 876662, "postDate": "2020-06-06T21:58:33.237Z", "content": "Thanks for your answer! Im still learning in terms of pooling methods etc.. :)
", "rawMarkdown": "Thanks for your answer! Im still learning in terms of pooling methods etc.. :)", "votes": 1 } ] }, { "id": 874989, "postDate": "2020-06-05T12:33:26.763Z", "content": "That also depends on how those 12/24 tiles are selected. If you use the same criterion, e.g. tiles selected according to number of pixels with tissue, your model will first be trained using images with, on average, more tissue than blank spaces, and if you then feed it images made of 24 tiles, the additional 12 will probably have much more white areas, which the model might be not so good at recognizing,
", "rawMarkdown": "That also depends on how those 12/24 tiles are selected. If you use the same criterion, e.g. tiles selected according to number of pixels with tissue, your model will first be trained using images with, on average, more tissue than blank spaces, and if you then feed it images made of 24 tiles, the additional 12 will probably have much more white areas, which the model might be not so good at recognizing,", "votes": 1 }, { "id": 874153, "postDate": "2020-06-04T17:02:38.507Z", "content": "Hi guys, I was just wondering, maybe someone of you could explain it to me:
\n\nWhat happens if we first train a multi instance learning model with 12 tiles. Later we decide to increase the number of tiles to 24. The backbone of the model should already have learned some features in my opinion from always using 12 tiles. Or will it be complete garbage and just rewrite the weights?
\n\nObviously, the main difference will be in the classifier, after the backbone.
", "rawMarkdown": "Hi guys, I was just wondering, maybe someone of you could explain it to me:\n\nWhat happens if we first train a multi instance learning model with 12 tiles. Later we decide to increase the number of tiles to 24. The backbone of the model should already have learned some features in my opinion from always using 12 tiles. Or will it be complete garbage and just rewrite the weights?\n\nObviously, the main difference will be in the classifier, after the backbone.", "votes": 2 } ], "comments": [ { "id": 875052, "author_name": "Arnaud Roussel", "author_url": "", "post_date": "2020-06-05T13:45:23.417000", "content": "This should be similar to doing progressive size learning where you start with a small image to kickstart your model and then use bigger images. It can lead to better performance for image classification so it might be worth looking into here also but by playing with the num_tiles instead of image size. Maybe more white areas can create issues though.
\n\nOn the other hand the major difference is that here you'd be first not giving all the information which could lead to higher overfit if the model learns irrelevant features when the cancer tiles are not included.
", "votes": 1, "replies": [ { "id": 875517, "author_name": "Claudio Fanconi", "author_url": "", "post_date": "2020-06-05T21:43:17.800000", "content": "Excellent answer, thank you very much!
\n\nThese are valid points. I also believe that a critical problem is that information was withheld and thus could have learned the wrong things.
", "votes": 0, "replies": [] }, { "id": 875556, "author_name": "Yann Majewski", "author_url": "", "post_date": "2020-06-05T22:55:29.917000", "content": "Is it possible to train a model with a variable number of tiles between the examples?
", "votes": 0, "replies": [] }, { "id": 876654, "author_name": "Arnaud Roussel", "author_url": "", "post_date": "2020-06-06T21:46:38.657000", "content": "@yannmajewski Only if you have a pooling function that can deal with the varrying length. Max pools, attention, LSTM, etc.
\n\nAverage pooling though cannot deal with it (in theory).
", "votes": 0, "replies": [] }, { "id": 876662, "author_name": "Yann Majewski", "author_url": "", "post_date": "2020-06-06T21:58:33.237000", "content": "Thanks for your answer! Im still learning in terms of pooling methods etc.. :)
", "votes": 1, "replies": [] } ] }, { "id": 874989, "author_name": "Marco Perini", "author_url": "", "post_date": "2020-06-05T12:33:26.763000", "content": "That also depends on how those 12/24 tiles are selected. If you use the same criterion, e.g. tiles selected according to number of pixels with tissue, your model will first be trained using images with, on average, more tissue than blank spaces, and if you then feed it images made of 24 tiles, the additional 12 will probably have much more white areas, which the model might be not so good at recognizing,
", "votes": 1, "replies": [] } ], "raw_markdown_by_id": { "875052": "This should be similar to doing progressive size learning where you start with a small image to kickstart your model and then use bigger images. It can lead to better performance for image classification so it might be worth looking into here also but by playing with the num_tiles instead of image size. Maybe more white areas can create issues though.\n\nOn the other hand the major difference is that here you'd be first not giving all the information which could lead to higher overfit if the model learns irrelevant features when the cancer tiles are not included.", "874989": "That also depends on how those 12/24 tiles are selected. If you use the same criterion, e.g. tiles selected according to number of pixels with tissue, your model will first be trained using images with, on average, more tissue than blank spaces, and if you then feed it images made of 24 tiles, the additional 12 will probably have much more white areas, which the model might be not so good at recognizing,", "874153": "Hi guys, I was just wondering, maybe someone of you could explain it to me:\n\nWhat happens if we first train a multi instance learning model with 12 tiles. Later we decide to increase the number of tiles to 24. The backbone of the model should already have learned some features in my opinion from always using 12 tiles. Or will it be complete garbage and just rewrite the weights?\n\nObviously, the main difference will be in the classifier, after the backbone." } }