Tweeted By @Reza_Zadeh

Neural nets are terrible at arithmetic & counting. If you train one in 1 to 10, it will do okay on 3 + 5 but fail miserably for 1000 + 3000. Resolving this,"Neural Arithmetic Logic Units" can track time, do arithmetic on images of numbers, & extrapolate https://t.co/15SSRpjMnR pic.twitter.com/Abw2qRnu5I

— Reza Zadeh (@Reza_Zadeh) August 17, 2018