3 ms·
Is this seriously done? I just completed a small literature review for neural program synthesis which uses inputs as training data to generate a program using n
by exdsq 7y ago
Is this seriously done? I just completed a small literature review for neural program synthesis which uses inputs as training data to generate a program using neural networks that can generate the right outputs but it's a huge search problem and I assume for anything non-trivial this is out of our reach? As an example, Deepmind achieved binary multiplication using this technique in 2018 and this was notable.
- goatinaboat 7y agoI’ve never done it but it should be feasible. An IP packet is 1518x8 == 12144 bits which is not especially big for a one-hot style encoding these days. You would just need to be sure to capture representative traffic e.g. accounting for monthly jobs.
- jerome-jh 7y agoIn my understanding, ML works by mapping a huge input space to a tiny output space, and most of the input space remains unexplored. In this case the input space is small and you want a quite dense mapping from input to output. I do not think ML applies.
- gdxhyrd 7y agoNo, ML works fine with large output spaces too. The problem is that the solution space grows super exponentially, and if you need to find an exact one, then the number of samples too.
- lonelappde 7y agoMost IO is larger than one packet.
- crusso 7y agoIs this seriously done? No. Not even close for any general purpose I/O.