5 ms·
RL is really really brittle, i.e. one simple bug that wouldn't matter in GAN/LSTM (such as e.g. adding one more or less number than needed) would cause your RL
by bitL 8y ago
RL is really really brittle, i.e. one simple bug that wouldn't matter in GAN/LSTM (such as e.g. adding one more or less number than needed) would cause your RL algorithm not to converge and you'd spend days-weeks debugging it - it requires extreme discipline verifying every single step of your algorithm. Often you don't know if your model didn't converge because it is not strong enough or if there was some tiny little error/wrong assumption leading to divergence.
- mark_l_watson 8y agoYes, some types of NNs are resilient to errors in implementation. I hired David Rumelhart as a consultant in the 1980s (when he was still at UCSD) when I was writing SAIC’s neural network product for Windows. He warned me that he was used to seeing projects with big implementation errors that still worked very well. I do find GANs brittle. I use them for generating numeric/categorical spreadsheet data, not for images.
- rbanffy 8y ago> He warned me that he was used to seeing projects with big implementation errors that still worked very well. For any sufficiently complex function, if you tune the constants enough, the wrong function will sometimes approximate the results the correct one would give if it were set up with the correct constants.
- avinium 8y ago> I do find GANs brittle. I use them for generating numeric/categorical spreadsheet data, not for images I’m interested to hear more about this. Care to elaborate?