{ "id": 11437, "title": "Any descriptor hand-designers around here ?", "url": "/competitions/datasciencebowl/discussion/11437", "author_name": "", "post_date": "2015-01-06T11:59:05.960Z", "votes": null, "comment_count": 2, "views": 1075, "content": "

Hi,

\n

I can understand after reading different posts that most competitors go for a neural network solution. 

\n

Is there anyone who is trying to build appropriate features ? Why should I prefer a neural network approach ? I have poor knowledge about neural network, but won't the solutions built by competitors using this approach resemble ?

\n

Thanks for sharing,

", "messages": [ { "id": "62339", "postDate": "01/06/2015 11:59:05", "content": "

Hi,

\n

I can understand after reading different posts that most competitors go for a neural network solution. 

\n

Is there anyone who is trying to build appropriate features ? Why should I prefer a neural network approach ? I have poor knowledge about neural network, but won't the solutions built by competitors using this approach resemble ?

\n

Thanks for sharing,

", "rawMarkdown": "", "votes": null }, { "id": "62381", "postDate": "01/06/2015 23:48:09", "content": "

It depends on what you are trying to achieve.  

\n

If your goal is to try to win the competition, it makes more sense to try to use:

\n\n

Especially if you can't be full time on the competition. 

\n

In the case of image classification, CNNs (Convolutional Neural Network) are now making very good results. Plus, it was a CNN which won the Galaxy Zoo challenge and if you look at the leaderboard right now sedielem is at the first position with his team. So if that's what they are doing, it might be good to check CNNs :) 

\n

But sometimes you might want to choose other path. For instance if you don't have a GPU or are very unfamiliar with the subject. We can't know for sure which model will win the competition, and that's the beauty of it. 

\n

If you are more thrilled by the research than by the prize, you might find it more interesting to try to improve a type of model you like, or create features of your own, to try to get some outstanding results rather than just running and tuning a model. However you should just be realistic with what to expect from the model your are using. Try to find papers talking about the state of the art performance using the solution. 

\n\n\n

I am not sure but I think JeffH is using something based on image dimensions and density

\n\n\n

Not really because you have a ton of things you can configure, the activation functions can be different, each person may use different features, you might do ensemble learning combining the results of several models. So they might look similar by the name of the model they use, but I believe each and every solution is unique. 

", "rawMarkdown": "", "votes": null }, { "id": "62383", "postDate": "01/07/2015 01:40:30", "content": "

Hi MPM,

\n\n

Alexandre has some good advice there. I am working with R and published some starter code that uses Random Forest on image geometry and very basic image calculations. I've added many other features like area of largest object, length of major/minor axis, stdev of intensity, and axis angles. My best model is now scoring 1.98 which seems pretty good to me based on simple measurements. I am interested in deep neural nets, but have not made progress with them yet.

\n\n

Good luck,

\n

Jeff

", "rawMarkdown": "", "votes": null } ], "comments": [ { "id": 62381, "author_name": "alexandrekowalczyk", "author_url": "", "post_date": "01/06/2015 23:48:09", "content": "

It depends on what you are trying to achieve.  

\n

If your goal is to try to win the competition, it makes more sense to try to use:

\n\n

Especially if you can't be full time on the competition. 

\n

In the case of image classification, CNNs (Convolutional Neural Network) are now making very good results. Plus, it was a CNN which won the Galaxy Zoo challenge and if you look at the leaderboard right now sedielem is at the first position with his team. So if that's what they are doing, it might be good to check CNNs :) 

\n

But sometimes you might want to choose other path. For instance if you don't have a GPU or are very unfamiliar with the subject. We can't know for sure which model will win the competition, and that's the beauty of it. 

\n

If you are more thrilled by the research than by the prize, you might find it more interesting to try to improve a type of model you like, or create features of your own, to try to get some outstanding results rather than just running and tuning a model. However you should just be realistic with what to expect from the model your are using. Try to find papers talking about the state of the art performance using the solution. 

\n\n\n

I am not sure but I think JeffH is using something based on image dimensions and density

\n\n\n

Not really because you have a ton of things you can configure, the activation functions can be different, each person may use different features, you might do ensemble learning combining the results of several models. So they might look similar by the name of the model they use, but I believe each and every solution is unique. 

", "votes": null, "replies": [] }, { "id": 62383, "author_name": "jeffhebert", "author_url": "", "post_date": "01/07/2015 01:40:30", "content": "

Hi MPM,

\n\n

Alexandre has some good advice there. I am working with R and published some starter code that uses Random Forest on image geometry and very basic image calculations. I've added many other features like area of largest object, length of major/minor axis, stdev of intensity, and axis angles. My best model is now scoring 1.98 which seems pretty good to me based on simple measurements. I am interested in deep neural nets, but have not made progress with them yet.

\n\n

Good luck,

\n

Jeff

", "votes": null, "replies": [] } ], "raw_markdown_by_id": { "62339": "", "62381": "", "62383": "" }, "source": "meta" }