Resultats MLP

Version 8, last updated by XavierMuller at April 30, 2010 08:26 UTC

Tous les meilleurs resultats pour le MLP peuvent etre trouve dans le tableau suivant.

Training Set	Fine Tuning Set	Training Set Validation Error	NIST Overall Validation Error (pre finetune)	NIST Overall Test Error (pre finetune)	NIST Overall Validation Error	NIST Overall Test Error	NIST Digit Test Error	NIST Character Test Error	NIST Majuscule Test Error	NIST Minuscule Test Error	Model Parameter	Model Path
NIST	None	12.27	12.27	24.19	12.27	24.19	3.45	14.19	11.45	16.20	nb hidden units =1000 initial learning rate=0.5 learning rate decay factor =0.5 nb exemples seen=4056480	xm_nist_final0_mlp/18
P07	None	53.00	20.16	24.32	20.16	24.32	4.85	19.62	15.44	21.32	nb hidden units =1000 learning rate =0.0075 tau=1e8 nb exemples seen=75075000	xvm_p07_lr2_refined_f1000/5
PNIST	None	40.58	17.62	22.81	17.62	22.81	3.85	14.71	10.68	16.68	nb hidden units =800 learning rate =0.0075 tau=1e8 nb exemples seen=77187500	xvm_pnist_lr2_refined_f0/5
07	NIST	53.00	20.16	24.32	13.10	23.31	3.26	13.42	10.81	15.80	nb hidden units =1000 learning rate =0.0075 tau=1e8 nb exemples seen=75075000 Fine-Tune: learning rate =0.1 tau=1e6 nb exemples seen=4062500	xvm_p07_postopt_lr2/3/
PNIST	NIST	40.58	17.62	22.81	12.98	23.13	3.18	13.14	10.39	14.92	nb hidden units =800 learning rate =0.0075 tau=1e8 nb exemples seen=77187500 Fine-Tune: learning rate =0.1 tau=1e6 nb exemples seen=4062500	xvm_pnist_postopt_lr2_1b/3

All the experiments where run with a softmax function at the output, a tanh non-linear function at the ouput of the hidden layer. The log-likelyhood error was used for the erro calculation. Mini-batches of 20 where used for the training.

The NIST digit error is calculated with the assumption that we know beforehand that the current exemple is a digit.

The NIST minuscule and majuscule error is calculated with the assumption that we know beforehand that the current exemple is either a minuscule or a majuscule.

The NIST character error is calculated with the assumption that we know beforehand that the current exemple is a character. We also do not distinguish between minsuscules and majuscules.