{ "id": 8233, "title": "Clarification about the rules", "url": "/competitions/seizure-detection/discussion/8233", "author_name": "", "post_date": "2014-05-20T15:22:04.323Z", "votes": 1, "comment_count": 2, "views": 1360, "content": "

[quote=Competition Rules]Classification must be performed by an algorithm. Visual review of data segments is not a permitted.[/quote]

\n

Does this section of the rules just imply that hand picked predictions are not allowed? The second sentence could be interpreted in more restrictive ways as well.

\n

[quote=Competition Rules]# Not allowed:
if 'Dog_1' then foo()
if 'Patient_1' then bar()
...

\n

# Allowed
if f(signal) < 2 then foo()
else bar()[/quote]

\n

Are we still allowed to have individual models for each subjects as long as they can be trained in automated way?

", "messages": [ { "id": "44952", "postDate": "05/20/2014 15:22:04", "content": "

[quote=Competition Rules]Classification must be performed by an algorithm. Visual review of data segments is not a permitted.[/quote]

\n

Does this section of the rules just imply that hand picked predictions are not allowed? The second sentence could be interpreted in more restrictive ways as well.

\n

[quote=Competition Rules]# Not allowed:
if 'Dog_1' then foo()
if 'Patient_1' then bar()
...

\n

# Allowed
if f(signal) < 2 then foo()
else bar()[/quote]

\n

Are we still allowed to have individual models for each subjects as long as they can be trained in automated way?

", "rawMarkdown": "", "votes": null }, { "id": "44955", "postDate": "05/20/2014 15:37:09", "content": "

As for question 1, basically we're saying we need you to generate an algorithm that does the classification. What we don't want, for example, is a trained epileptologist looking through the data saying yes or no, and then the "algorithm" submitted becomes a large switch statement with the expert review results. You certainly are allowed plot and review the data as you develop your algorithm though. Reviewing raw data, frequency spectra, and other candidate features is allowed as long as your submitted classifications are generated by an algorithm.

\n\n

You're also allowed to retrain algorithms on each subject. However, what is not allowed is to have a completely new algorithm or set of features for each subject. As the sample code shows- it's OK to have flow control statements based on features you calculate from the data, but not OK to key on the subject's ID and run a totally separate process for each subject.

", "rawMarkdown": "", "votes": null }, { "id": "50203", "postDate": "07/09/2014 14:13:17", "content": "

Thank you for the question to Herra Hu and the clarification to bbrinkm.

\n\n

So, let me know if I am mistaken, you can ( and probably must) train a model for each subject but you cannot change that model depending on the name of the files, sampling frequency as said in this thread. 

\n\n

For example we can  :

\n\n

- Have all the data together and train a unique model.

\n\n

- Distinguish among subjects and have a classifier trained for each subject, but all this classifiers are of the same type.

\n\n

- Extract some feature from the signal (variable data in .mat files) to select among different classifiers. For example have one for dogs ( which could have a signal such and such) and humans  ( which could show a difference in some feature of the signal). 

\n\n

But we cannot:

\n\n

- Have a classifier trained with all the dogs and a different one with all the patients, using to distinguish the fact that the name of the file for dog is Dog_... and for patient is Patient_

\n

- Use the frequency to separate Dogs and Patients.

\n\n

Thanks in advance.

\n\n

Jesús

", "rawMarkdown": "", "votes": null } ], "comments": [ { "id": 44955, "author_name": "bbrinkm", "author_url": "", "post_date": "05/20/2014 15:37:09", "content": "

As for question 1, basically we're saying we need you to generate an algorithm that does the classification. What we don't want, for example, is a trained epileptologist looking through the data saying yes or no, and then the "algorithm" submitted becomes a large switch statement with the expert review results. You certainly are allowed plot and review the data as you develop your algorithm though. Reviewing raw data, frequency spectra, and other candidate features is allowed as long as your submitted classifications are generated by an algorithm.

\n\n

You're also allowed to retrain algorithms on each subject. However, what is not allowed is to have a completely new algorithm or set of features for each subject. As the sample code shows- it's OK to have flow control statements based on features you calculate from the data, but not OK to key on the subject's ID and run a totally separate process for each subject.

", "votes": null, "replies": [] }, { "id": 50203, "author_name": "jesusfbes", "author_url": "", "post_date": "07/09/2014 14:13:17", "content": "

Thank you for the question to Herra Hu and the clarification to bbrinkm.

\n\n

So, let me know if I am mistaken, you can ( and probably must) train a model for each subject but you cannot change that model depending on the name of the files, sampling frequency as said in this thread. 

\n\n

For example we can  :

\n\n

- Have all the data together and train a unique model.

\n\n

- Distinguish among subjects and have a classifier trained for each subject, but all this classifiers are of the same type.

\n\n

- Extract some feature from the signal (variable data in .mat files) to select among different classifiers. For example have one for dogs ( which could have a signal such and such) and humans  ( which could show a difference in some feature of the signal). 

\n\n

But we cannot:

\n\n

- Have a classifier trained with all the dogs and a different one with all the patients, using to distinguish the fact that the name of the file for dog is Dog_... and for patient is Patient_

\n

- Use the frequency to separate Dogs and Patients.

\n\n

Thanks in advance.

\n\n

Jesús

", "votes": null, "replies": [] } ], "raw_markdown_by_id": { "44952": "", "44955": "", "50203": "" }, "source": "meta" }