We found a result of 91% similarity relative to non-memristor tests. Our R-tolerance was 1% (we think). We found a true accuracy of 82% (relative to correct results) in our memristor model, same as the researchers.
TRUE ACCURACY:
testresult_withoutmems 83.55
testresult_withmems01 82.00
testresult_withmems1 80.70
testresult_withmems2 52.05
testresult_withmems3 12.40
Difference between without and with mems: 1.55%
HOW IDENTICAL TO NONMEMS:
testresult_withoutmems 100.00
testresult_withmems01 91.25
testresult_withmems1 87.30
testresult_withmems2 52.15
testresult_withmems3 11.95
R tolerance - basically, how close the trial’s resistance must be to predicted resistance for the update process to complete. If R tolerance is low, trials last longer and must run more updates to work. If R tolerance is high, trials take less time but may be more inaccurate.
https://github.com/hjq310/neuropack
NeuroPack: An Algorithm-level Python-based Simulator for...
We tried using the simulation for 100 runs:
