My Machine Learning Idea – You Can Use A Parity Method To Check Classification Errors. Possibly Correcting Single Errors Given Three Or So Parity Bits. Testing On MNIST

Basically you encode the labels with the error correcting fomula. Like the Hamming Code. Just google ecc error correcting code.

Then train. On the test set you can then verify a little that the classification is correct by computing the three parity bits from the data and compare. It should also be possible to correct a classification.

Here I got a data bit error [1]. 4 bits index [0,1,2,4] were data bits. The other 3 were parity bits.

The correct was y_test[1].