petr-kovalev/punched-cards-oracle-mnist

Object recognition by random binary data lookup for Oracle MNIST

'Punched cards' for Oracle-MNIST

Object recognition by sparse random binary data lookup. Based on this article

Performing single-shot Oracle-MNIST ancient characters recognition by lookup over the most representative sparse input bit sets of the training data (out of 28⋅28⋅8 = 6272 bits per training sample)

Bit vector set similarity evaluation using the maximum spanning tree is described in this article

The same algorithm applied to the QMNIST dataset is here

The same algorithm applied to the Fashion-MNIST dataset is here

Program output:

Punched card bit length: 8

Average single-shot correct recognitions on fine-tune iteration: 4580, 4560

Top punched card per input:
Training results: 4702 correct recognitions of 27222
Test results: 498 correct recognitions of 3000

Top 39 (5%) punched cards per input:
Training results: 7181 correct recognitions of 27222
Test results: 739 correct recognitions of 3000

All punched cards:
Training results: 8397 correct recognitions of 27222
Test results: 792 correct recognitions of 3000

Punched card bit length: 16

Average single-shot correct recognitions on fine-tune iteration: 4814, 4836, 4836, 4832

Top punched card per input:
Training results: 5624 correct recognitions of 27222
Test results: 568 correct recognitions of 3000

Top 19 (5%) punched cards per input:
Training results: 8346 correct recognitions of 27222
Test results: 844 correct recognitions of 3000

All punched cards:
Training results: 9687 correct recognitions of 27222
Test results: 947 correct recognitions of 3000

Punched card bit length: 32

Average single-shot correct recognitions on fine-tune iteration: 5295, 5384, 5418, 5434, 5443, 5447, 5446

Top punched card per input:
Training results: 6741 correct recognitions of 27222
Test results: 743 correct recognitions of 3000

Top 9 (5%) punched cards per input:
Training results: 9077 correct recognitions of 27222
Test results: 974 correct recognitions of 3000

All punched cards:
Training results: 10854 correct recognitions of 27222
Test results: 1065 correct recognitions of 3000

Punched card bit length: 64

Average single-shot correct recognitions on fine-tune iteration: 5883, 6030, 6090, 6117, 6135, 6146, 6152, 6154, 6156, 6155

Top punched card per input:
Training results: 7265 correct recognitions of 27222
Test results: 774 correct recognitions of 3000

Top 4 (5%) punched cards per input:
Training results: 8546 correct recognitions of 27222
Test results: 900 correct recognitions of 3000

All punched cards:
Training results: 11509 correct recognitions of 27222
Test results: 1111 correct recognitions of 3000

Punched card bit length: 128

Average single-shot correct recognitions on fine-tune iteration: 6568, 6711, 6777, 6818, 6841, 6858, 6866, 6877, 6882, 6886, 6887, 6891, 6890

Top punched card per input:
Training results: 7803 correct recognitions of 27222
Test results: 812 correct recognitions of 3000

Top 2 (5%) punched cards per input:
Training results: 8493 correct recognitions of 27222
Test results: 895 correct recognitions of 3000

All punched cards:
Training results: 11955 correct recognitions of 27222
Test results: 1168 correct recognitions of 3000

Punched card bit length: 256

Average single-shot correct recognitions on fine-tune iteration: 7408, 7517, 7567, 7608, 7631, 7650, 7662, 7672, 7682, 7688, 7696, 7699, 7699, 7705, 7704

Top punched card per input:
Training results: 8384 correct recognitions of 27222
Test results: 876 correct recognitions of 3000

Top 1 (5%) punched cards per input:
Training results: 8384 correct recognitions of 27222
Test results: 876 correct recognitions of 3000

All punched cards:
Training results: 11949 correct recognitions of 27222
Test results: 1163 correct recognitions of 3000

Press "Enter" to exit the program...