scaled_fa update: s beomces zero #12
Comments
|
Well if s is being set to zero then this means that the likelihood is being maximized at s = 0? Or is this a bug? I would verify this first. If it is correct—not a bug—then this means that this mixture component is behaving as a delta mass. I'm not sure how best to deal with this case, but it seems that the simplest approach would be to constrain s to be positive. We could potentially accommodate s = 0 but it would make the code significantly more complicated. |
|
i think s=0 is a valid mle, so we should try to allow for it.
…On Wed, Dec 8, 2021 at 8:58 AM Peter Carbonetto ***@***.***> wrote:
Well if s is being set to zero then this means that the likelihood is
being maximized at s = 0? Or is this a bug? I would verify this first. If
it is correct—not a bug—then this means that this mixture component is
behaving as a delta mass. I'm not sure how best to deal with this case, but
it seems that the simplest approach would be to constrain s to be positive.
We could potentially accommodate s = 0 but it would make the code
significantly more complicated.
—
You are receiving this because you are subscribed to this thread.
Reply to this email directly, view it on GitHub
<#12 (comment)>, or
unsubscribe
<https://github.com/notifications/unsubscribe-auth/AANXRRMTXNQWUZAQP652643UP5XCXANCNFSM5JSH6UEQ>
.
Triage notifications on the go with GitHub Mobile for iOS
<https://apps.apple.com/app/apple-store/id1477376905?ct=notification-email&mt=8&pt=524675>
or Android
<https://play.google.com/store/apps/details?id=com.github.android&referrer=utm_campaign%3Dnotification-email%26utm_medium%3Demail%26utm_source%3Dgithub>.
|
|
I'm not sure why NA is being obtained - is this an issue with optim?
If so, can we fix it by just defining the update to "do nothing" if
initialized at s=0?
On Wed, Dec 8, 2021 at 11:12 AM Matthew Stephens ***@***.***>
wrote:
… i think s=0 is a valid mle, so we should try to allow for it.
On Wed, Dec 8, 2021 at 8:58 AM Peter Carbonetto ***@***.***>
wrote:
> Well if s is being set to zero then this means that the likelihood is
> being maximized at s = 0? Or is this a bug? I would verify this first. If
> it is correct—not a bug—then this means that this mixture component is
> behaving as a delta mass. I'm not sure how best to deal with this case, but
> it seems that the simplest approach would be to constrain s to be positive.
> We could potentially accommodate s = 0 but it would make the code
> significantly more complicated.
>
> —
> You are receiving this because you are subscribed to this thread.
> Reply to this email directly, view it on GitHub
> <#12 (comment)>,
> or unsubscribe
> <https://github.com/notifications/unsubscribe-auth/AANXRRMTXNQWUZAQP652643UP5XCXANCNFSM5JSH6UEQ>
> .
> Triage notifications on the go with GitHub Mobile for iOS
> <https://apps.apple.com/app/apple-store/id1477376905?ct=notification-email&mt=8&pt=524675>
> or Android
> <https://play.google.com/store/apps/details?id=com.github.android&referrer=utm_campaign%3Dnotification-email%26utm_medium%3Demail%26utm_source%3Dgithub>.
>
>
|
|
oh, i see, it is in the FA update, not optim.
Yes, let's just define the FA update to keep s=0 if initialized at s=0.
On Wed, Dec 8, 2021 at 11:14 AM Matthew Stephens ***@***.***>
wrote:
… I'm not sure why NA is being obtained - is this an issue with optim?
If so, can we fix it by just defining the update to "do nothing" if
initialized at s=0?
On Wed, Dec 8, 2021 at 11:12 AM Matthew Stephens ***@***.***>
wrote:
> i think s=0 is a valid mle, so we should try to allow for it.
>
> On Wed, Dec 8, 2021 at 8:58 AM Peter Carbonetto ***@***.***>
> wrote:
>
>> Well if s is being set to zero then this means that the likelihood is
>> being maximized at s = 0? Or is this a bug? I would verify this first. If
>> it is correct—not a bug—then this means that this mixture component is
>> behaving as a delta mass. I'm not sure how best to deal with this case, but
>> it seems that the simplest approach would be to constrain s to be positive.
>> We could potentially accommodate s = 0 but it would make the code
>> significantly more complicated.
>>
>> —
>> You are receiving this because you are subscribed to this thread.
>> Reply to this email directly, view it on GitHub
>> <#12 (comment)>,
>> or unsubscribe
>> <https://github.com/notifications/unsubscribe-auth/AANXRRMTXNQWUZAQP652643UP5XCXANCNFSM5JSH6UEQ>
>> .
>> Triage notifications on the go with GitHub Mobile for iOS
>> <https://apps.apple.com/app/apple-store/id1477376905?ct=notification-email&mt=8&pt=524675>
>> or Android
>> <https://play.google.com/store/apps/details?id=com.github.android&referrer=utm_campaign%3Dnotification-email%26utm_medium%3Demail%26utm_source%3Dgithub>.
>>
>>
>
|
|
Okay, then in some parts of the code an if-else statement will be needed check for the special case that the prior covariance matrix U is a matrix of zeros. |
|
I added a check to fa_scaled_iid() and fa_scaled_notiid() to check whether s==0 or U0 == a matrix of all 0s. And I checked on simple simulated data by input a matrix of all 0s, it works fine. |
|
What happens to the posterior probabilities computed in |
|
In the real data example, s = 0 component has exactly 0 weight. While for the matrix of all 0s, the weight is very close to 0, 3.250872e-317. But I don't think there is a relationship between s=0/U all zeros and weight being 0? |
|
I agree i don't see that there would be an obvious relationship
…On Wed, Dec 8, 2021, 19:42 yunqiyang0215 ***@***.***> wrote:
In the real data example, s = 0 component has exactly 0 weight. While for
the matrix of all 0s, the weight is very close to 0, 3.250872e-317. But I
don't think there is a relationship between s=0/U all zeros and weight
being 0?
—
You are receiving this because you commented.
Reply to this email directly, view it on GitHub
<#12 (comment)>, or
unsubscribe
<https://github.com/notifications/unsubscribe-auth/AANXRRJHTJQDMHQKJQ4A6JLUQACR7ANCNFSM5JSH6UEQ>
.
Triage notifications on the go with GitHub Mobile for iOS
<https://apps.apple.com/app/apple-store/id1477376905?ct=notification-email&mt=8&pt=524675>
or Android
<https://play.google.com/store/apps/details?id=com.github.android&referrer=utm_campaign%3Dnotification-email%26utm_medium%3Demail%26utm_source%3Dgithub>.
|
|
Sorry that's an error on my part. It does make me think that for some of the U updates we might want to add checks for U being a matrix of zeros. For example, if U is a matrix of zeros, will the ED update have any effect? |
|
i think it won't have any effect
…On Thu, Dec 9, 2021 at 10:44 AM Peter Carbonetto ***@***.***> wrote:
Sorry that's an error on my part.
It does make me think that for some of the U updates we might want to add
checks for U being a matrix of zeros. For example, if U is a matrix of
zeros, will the ED update have any effect?
—
You are receiving this because you commented.
Reply to this email directly, view it on GitHub
<#12 (comment)>, or
unsubscribe
<https://github.com/notifications/unsubscribe-auth/AANXRRIAVTHRUG2754GWCQ3UQDMHDANCNFSM5JSH6UEQ>
.
Triage notifications on the go with GitHub Mobile for iOS
<https://apps.apple.com/app/apple-store/id1477376905?ct=notification-email&mt=8&pt=524675>
or Android
<https://play.google.com/store/apps/details?id=com.github.android&referrer=utm_campaign%3Dnotification-email%26utm_medium%3Demail%26utm_source%3Dgithub>.
|
|
Yeah so we might want to give a warning or error for this case since it could conceivably happen. |
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
I noticed that from the ukbblood cell data example, an error arises possibly due to the fact the updated s becomes 0 in the current iteration. Then in the next iteration, estimated s will be NA. In the previous iteration, s=0.002099687.
The bug should occur in fa_scaled_iid() and fa_scaled_notiid(). The input s shouldn't be 0. One way is to add a constraint to s before running fa_scaled_iid().
However, if input s=0, fa_scaled_iid() couldn't run. Not sure how s=NA is created.
The text was updated successfully, but these errors were encountered: