Bonjour à tous (English will follow),
j'ai mis à jour mes diapositives sur les réseaux à convolution:
1) J'étais pas super content avec ma description du concept de convolution. J'ai donc ajouté un autre exemple, étape par étape, de l'application d'une convolution, ainsi qu'une description un peu plus explicite de la relation entre l'opération de convolution et le calcul de l'activation linéaire d'une "feature map"
2) J'ai ajouté la formule du gradient p/r aux poids d'une feature map, dans l'opération de convolution
3) J'ai corrigé le gradient p/r à l'entrée du "average pooling / subsampling". Il y avait une erreur dans la formule que j'avais donné (10 Hugo points pour moi, pour avoir trouvé l'erreur ;-) ).
===
Hi all,
I updated my slides on convolutional networks:
1) I was not super happy with my description of the concept of convolution. So I added another example, step by step, of applying a convolution and added a more explicit description of the relationship between the convolution operation and the calculation of the linear activation of a "feature map"
2) I added the gradient formula w/r to the weights of a feature map in the convolution operation
3) I fixed the gradient w/r to the input of the "average pooling / subsampling". There was an error in the formula I gave (10 Hugo points for me, for finding the error ;-)).