Of course you could train a NN to do arithmetic, but this is much more impressive. Training a NN network to solve problems with available tools means more abstraction, and is closer to AGI than just essentially learning a LUT.

> Of course you could train a NN to do arithmetic

Are we really capable of teaching a NN to parse and calculate an arbitrary arithmetic expression? Because that sounds incredibly impressive...

Yes, it can be done.

https://openreview.net/pdf?id=S1eZYeHFDS

Natural language is harder.

I'm not sure. Human is general intelligence, but has to learn basic maths too.

Here’s one that does so using roman numerals:

http://static.offd.es/numerals/

It’s unsurprisingly easy to implement

Interesting but can it do more than addition? Also it doesn't seems to have 100% accuracy.

Yeah. I only trained addition. Actually exploring the impact of training a net to perform a range of operations on the minimum plausible neuron count would be quite interesting

I don’t see any reason why it would be significantly harder to do, however

You’re right about accuracy. I didn’t let the model train enough to push the error low enough to guarantee exact results over the input range. But then again this was designed as a toy experiment, not something people should rely on