Skip to main content

MLP for MultiMNIST

Created on October 26|Last edited on October 26

Best results


Showing first 10 runs
05101520Step0.30.350.40.450.50.55
Showing first 10 runs
05101520Step0.40.50.60.70.80.9
Showing first 10 runs
05101520Step0.020.040.060.08
Showing first 10 runs
05101520Step0.60.811.21.41.61.8
Showing first 10 runs
05101520Step0.70.720.740.760.780.80.820.84
Showing first 10 runs
05101520Step0.10.20.30.40.50.6
Run set
396


Different accuracy metrics used:

There are 3 accuracy metrics used. In 1st, we give +0.5 if there is 1 correct guess in top2 predicted by model. And +1 for both correct guesses. We can use it as we know dataset has always 2 digits.
2nd one is Hamming similarity.
3rd one is complete match. This accuracy is lowest, but is actually pretty high for CNN. It is a good candidate for actual accuracy.


Effect of layers;


Run set
396

  • 0 number of layers, (so something similar to logistic regression gives best results).

Comparison:


Run set
5

  • We see that best accuracy comes from < 2 number of layers, and low lr.