002-Experiments with ENIST, Logistic regression

1.
Setting​

•
Normalize data to (0-0.25), Gaussian blur, to simulate spectrum addition​

•
channel to a single value, which is then followed by a linear layer.​

•
Logistic regression, BCE loss.​

2.
Experiment 1: MNIST split​

2.1 Config

•
MNIST data, official split to training and test. ​

•
0-7 classes as the recognition target. ​

•
Full CNN, 3 CNN layers, LayerNorm+ReLU[activation red]+Maxpooling. The last CNN layer compress each ​

•
Check EER by pooling all the positive and negative recognition of the 8 classes​

•
Compute the BCE of the positive samples (target as 1) and the negative samples (target as 0)​

•
x: image; y: label; p1=f(w1*x1) p2=f(w2*x2) p3=f(w1*x1+w2*x2); ​

•
Training: weigh w sampling from 0.2-1.0; mixture w1 and w2 sampling from 0.2-1.0; w1+w2=1.0; ​

•
Test: sampling weight and mixing 0-3 images from class 0-7​

2.2 Main results

Positive/Negative BCE

002-Experiments with ENIST, Logistic regression​