{ "id": 10449, "title": "Human specific model", "url": "/competitions/seizure-prediction/discussion/10449", "author_name": "", "post_date": "2014-09-25T15:10:07.837Z", "votes": 10, "comment_count": 5, "views": 2218, "content": "

I understand from the rules, that changing the model or the parameters based on direct knowledge of the subject is not allowed. But can we at least use the knowledge of whether the subject is a Human or a Dog?

\n

We can also infer it anyway based on heuristics or CV based decisions but I would like to know if we can use the information directly.

", "messages": [ { "id": "55230", "postDate": "09/25/2014 15:10:07", "content": "

I understand from the rules, that changing the model or the parameters based on direct knowledge of the subject is not allowed. But can we at least use the knowledge of whether the subject is a Human or a Dog?

\n

We can also infer it anyway based on heuristics or CV based decisions but I would like to know if we can use the information directly.

", "rawMarkdown": "", "votes": null }, { "id": "55237", "postDate": "09/25/2014 16:27:20", "content": "

I also think that this rule is incoherent, and was disappointed to see it carried over unchanged after the discussion during the previous competition. 

\n

I would like to see a simplification to something like "Anything except human labeling of test data is allowed". 

", "rawMarkdown": "", "votes": null }, { "id": "55321", "postDate": "09/27/2014 01:16:32", "content": "

Is the following construction allowed? 

\n

If (Number_of_electrodes ==N) then model1

\n

        else model2

", "rawMarkdown": "", "votes": null }, { "id": "55655", "postDate": "10/05/2014 21:39:14", "content": "

This is made clear in the rules:

\n\n

Participants must use an algorithm to classify the data segments. Any changes to the methodology across species/subjects must be done in an automated way, so that your approach will generalize to new subjects. Training algorithms separately on individual subjects is allowed. For example:

\n\n

# Not allowed:
if 'Dog_1' then foo()
if 'Patient_1' then bar()
...

\n

# Allowed
if f(signal) < 2 then foo()
else bar()

\n

#Also Allowed
for subject=1:N
   train(subject)
   test(subject)

\n\n

Basically, I read this as do as you please, but always generalize and automate.

", "rawMarkdown": "", "votes": null }, { "id": "56258", "postDate": "10/19/2014 16:25:01", "content": "

Could you please clarify if it is allowed to have different models  for each subject? For example, based on  validation results for dog1 (or patient)  the best neural network has 20 hidden units, for dog2 it has 40 units. And the same with features: some features work better for one subject and worse for another.  Parameters  and feature selection can be done in automated way, but I am still not sure if  it doesn't violate the rules. Thank you very much!

", "rawMarkdown": "", "votes": null }, { "id": "56728", "postDate": "10/25/2014 11:08:08", "content": "

Looking at the rules I guess that would not be allowed. You can't alter the model *specifically* for each patient but you can alter the way the software processes according to variances in the incoming data. This way the software should work for patients that are currently unseen to us.

", "rawMarkdown": "", "votes": null } ], "comments": [ { "id": 55237, "author_name": "emolson", "author_url": "", "post_date": "09/25/2014 16:27:20", "content": "

I also think that this rule is incoherent, and was disappointed to see it carried over unchanged after the discussion during the previous competition. 

\n

I would like to see a simplification to something like "Anything except human labeling of test data is allowed". 

", "votes": null, "replies": [] }, { "id": 55321, "author_name": "mkozine", "author_url": "", "post_date": "09/27/2014 01:16:32", "content": "

Is the following construction allowed? 

\n

If (Number_of_electrodes ==N) then model1

\n

        else model2

", "votes": null, "replies": [] }, { "id": 55655, "author_name": "kevinjos", "author_url": "", "post_date": "10/05/2014 21:39:14", "content": "

This is made clear in the rules:

\n\n

Participants must use an algorithm to classify the data segments. Any changes to the methodology across species/subjects must be done in an automated way, so that your approach will generalize to new subjects. Training algorithms separately on individual subjects is allowed. For example:

\n\n

# Not allowed:
if 'Dog_1' then foo()
if 'Patient_1' then bar()
...

\n

# Allowed
if f(signal) < 2 then foo()
else bar()

\n

#Also Allowed
for subject=1:N
   train(subject)
   test(subject)

\n\n

Basically, I read this as do as you please, but always generalize and automate.

", "votes": null, "replies": [] }, { "id": 56258, "author_name": "golondrina", "author_url": "", "post_date": "10/19/2014 16:25:01", "content": "

Could you please clarify if it is allowed to have different models  for each subject? For example, based on  validation results for dog1 (or patient)  the best neural network has 20 hidden units, for dog2 it has 40 units. And the same with features: some features work better for one subject and worse for another.  Parameters  and feature selection can be done in automated way, but I am still not sure if  it doesn't violate the rules. Thank you very much!

", "votes": null, "replies": [] }, { "id": 56728, "author_name": "davidjames", "author_url": "", "post_date": "10/25/2014 11:08:08", "content": "

Looking at the rules I guess that would not be allowed. You can't alter the model *specifically* for each patient but you can alter the way the software processes according to variances in the incoming data. This way the software should work for patients that are currently unseen to us.

", "votes": null, "replies": [] } ], "raw_markdown_by_id": { "55230": "", "55237": "", "55321": "", "55655": "", "56258": "", "56728": "" }, "source": "meta" }