4 ms·
My personal hope is that this type of development will lead to the solution of one of the biggest problems with neural networks: their intransparency. If the r
by vog 9y ago
My personal hope is that this type of development will lead to the solution of one of the biggest problems with neural networks: their intransparency.
If the result of the learning and optimization is not just a bunch of weights in a graph, but a readable algorithm, that would be a huge step forward into seeing what was really learned. Also, the results might be more stable against the usual Deep Learning attacks.
Of course, the result would read more like disassembled machine code, but that's still better than what we had before. Human tasks might then include finding good variable names, rearranging the code for clarity, etc. That is, typical reverse engineering work.
- andrewflnr 9y agoThis perspective strikes me as far too optimistic, for mostly the reasons you describe. I don't think evolving a general purpose programming language (even if not Turing-complete) is likely to produce more comprehensible systems than evolving a matrix, and in fact I would expect the opposite. Reverse engineering benefits from knowing that the original author was a human that can only handle so much complexity at once. Reverse engineering this sort of code will be more like molecular biology, where everything still depends on everything but you don't know how anymore, as opposed to linear algebra where the relationships are at least regular.