Version 1, last updated by humel at April 22, 2010 19:58 UTC

Objectif

  1. Illustrate experimentaly the advantage of convolutional networks for image recognition over traditional deep MLP (with or without unsupervised pretraining).
  2. Benchmark the model on the NIST dataset using the results obtained by Sylvain Pannetier to select the best training strategy

 

Experiences

SETUP

 

Convolutional Layers ([nb_filters,x_size,y_size]) ([52,5,5], [32,3,3]) , ([52,7,7], [52,3,3])
number hidden units (MLP) (1000),(500)
Maxpool
([2,2],[2,2])
Corruption level ([0.2,0.1])
Minibatch size (100)
Number epoch pretrain /layer (10)
Pretrain learning rate (0.01)
Finetune learning rate (0.1),(0.01)
Pretraining Dataset (P07)
Finteuning Dataset (P07)

Results

Experiences currently running.