{ "id": 176933, "title": "Linear decay as opposed to quantile regression", "url": "/competitions/osic-pulmonary-fibrosis-progression/discussion/176933", "author_name": "", "post_date": "2020-08-24T06:26:06.656777Z", "votes": 4, "comment_count": 5, "views": 0, "content": "

Sorry for the question, but what are the comparative benefits of a liner-decay based model over a quantile-regression based model?

", "messages": [ { "id": "983262", "postDate": "08/24/2020 06:26:06", "content": "

Sorry for the question, but what are the comparative benefits of a liner-decay based model over a quantile-regression based model?

", "rawMarkdown": "Sorry for the question, but what are the comparative benefits of a liner-decay based model over a quantile-regression based model?", "votes": null }, { "id": "983324", "postDate": "08/24/2020 07:24:22", "content": "

Hi @nxrprime can you elaborate of linear decay? thks in advance

", "rawMarkdown": "Hi @nxrprime can you elaborate of linear decay? thks in advance", "votes": null }, { "id": "983325", "postDate": "08/24/2020 07:25:18", "content": "

Linear decay as enumerated in this brilliant notebook: https://www.kaggle.com/miklgr500/linear-decay-based-on-resnet-cnn

", "rawMarkdown": "Linear decay as enumerated in this brilliant notebook: https://www.kaggle.com/miklgr500/linear-decay-based-on-resnet-cnn", "votes": null }, { "id": "983392", "postDate": "08/24/2020 09:13:33", "content": "

Good method to not overfit the model :)

", "rawMarkdown": "Good method to not overfit the model :)", "votes": null }, { "id": "983459", "postDate": "08/24/2020 10:20:03", "content": "

Giving the model some prior 'knowledge' - here that FVC decays linearly with time - is probably helpful if the knowledge is correct.

\n

Neural network with quantile loss could learn from the data that FVC decays linearly (if that's what it does) but would have to learn it from noisy data.

\n

I find it helpful to think of model choice as imposing some assumptions e.g. a linear regression imposes the assumption that variables interact linearly to produce the dependent variable.

\n

And the 'right' model for the data is the one that fits the out of sample data best.

", "rawMarkdown": "Giving the model some prior 'knowledge' - here that FVC decays linearly with time - is probably helpful if the knowledge is correct.\n\nNeural network with quantile loss could learn from the data that FVC decays linearly (if that's what it does) but would have to learn it from noisy data.\n\nI find it helpful to think of model choice as imposing some assumptions e.g. a linear regression imposes the assumption that variables interact linearly to produce the dependent variable. \n\nAnd the 'right' model for the data is the one that fits the out of sample data best.", "votes": null }, { "id": "1004973", "postDate": "09/10/2020 07:12:42", "content": "

Could someone elaborate the idea of choosing a random image ? This is also done during inference ? I'm a bit confused

", "rawMarkdown": "Could someone elaborate the idea of choosing a random image ? This is also done during inference ? I'm a bit confused", "votes": null } ], "comments": [ { "id": 983324, "author_name": "ulrich07", "author_url": "", "post_date": "08/24/2020 07:24:22", "content": "

Hi @nxrprime can you elaborate of linear decay? thks in advance

", "votes": null, "replies": [ { "id": 983325, "author_name": "nxrprime", "author_url": "", "post_date": "08/24/2020 07:25:18", "content": "

Linear decay as enumerated in this brilliant notebook: https://www.kaggle.com/miklgr500/linear-decay-based-on-resnet-cnn

", "votes": null, "replies": [] } ] }, { "id": 983392, "author_name": "eliasgreen", "author_url": "", "post_date": "08/24/2020 09:13:33", "content": "

Good method to not overfit the model :)

", "votes": null, "replies": [] }, { "id": 983459, "author_name": "jameschapman19", "author_url": "", "post_date": "08/24/2020 10:20:03", "content": "

Giving the model some prior 'knowledge' - here that FVC decays linearly with time - is probably helpful if the knowledge is correct.

\n

Neural network with quantile loss could learn from the data that FVC decays linearly (if that's what it does) but would have to learn it from noisy data.

\n

I find it helpful to think of model choice as imposing some assumptions e.g. a linear regression imposes the assumption that variables interact linearly to produce the dependent variable.

\n

And the 'right' model for the data is the one that fits the out of sample data best.

", "votes": null, "replies": [] }, { "id": 1004973, "author_name": "alexj21", "author_url": "", "post_date": "09/10/2020 07:12:42", "content": "

Could someone elaborate the idea of choosing a random image ? This is also done during inference ? I'm a bit confused

", "votes": null, "replies": [] } ], "raw_markdown_by_id": { "983262": "Sorry for the question, but what are the comparative benefits of a liner-decay based model over a quantile-regression based model?", "983324": "Hi @nxrprime can you elaborate of linear decay? thks in advance", "983325": "Linear decay as enumerated in this brilliant notebook: https://www.kaggle.com/miklgr500/linear-decay-based-on-resnet-cnn", "983392": "Good method to not overfit the model :)", "983459": "Giving the model some prior 'knowledge' - here that FVC decays linearly with time - is probably helpful if the knowledge is correct.\n\nNeural network with quantile loss could learn from the data that FVC decays linearly (if that's what it does) but would have to learn it from noisy data.\n\nI find it helpful to think of model choice as imposing some assumptions e.g. a linear regression imposes the assumption that variables interact linearly to produce the dependent variable. \n\nAnd the 'right' model for the data is the one that fits the out of sample data best.", "1004973": "Could someone elaborate the idea of choosing a random image ? This is also done during inference ? I'm a bit confused" }, "source": "meta" }