[Kaeldric]

Hello.

I'm studying the GoogLeNet by implementing a personal implementation in Tensorflow.
Last day I introduced a bug and changed the structure of the inception block a little.
Before that the algorithm run very poorly ... actually it never learned anything. 
After the modification (that changed the structure from the official paper one) the network start working as a charm  (with a little overfitting, but this is another story)!

You can find the two different structures here ---> https://imgur.com/a/vQiy0Te
where the bad one is the inception block described in the original paper and the good one is the modified version.

Is there a way to know why this modification leaded to a such big performance improvement? 

Thank you. 

[spydaz]

Hello.

I'm studying the GoogLeNet by implementing a personal implementation in Tensorflow.
Last day I introduced a bug and changed the structure of the inception block a little.
Before that the algorithm run very poorly ... actually it never learned anything. 
After the modification (that changed the structure from the official paper one) the network start working as a charm  (with a little overfitting, but this is another story)!

You can find the two different structures here ---> https://imgur.com/a/vQiy0Te
where the bad one is the inception block described in the original paper and the good one is the modified version.

Is there a way to know why this modification leaded to a such big performance improvement? 

Thank you. 

In this case the picture does not tell a thousand tales .... the pictures do not have enough information to make any sound judgement...at a glance its obvious that optimising the structure has made a difference although immeasurable but the information contained in the picture

[Kaeldric]

In short .... there is no way to know. 

[spydaz]

In short .... there is no way to know. 

So a futile question