Re:Computing glitch may have doomed Mars lander
by hliyan on Tuesday October 25, 2016 @07:31AM
(#12786969)
I was about to suggest n-version programming redundancy, when it occurred to me: aren't we starting to reach a point where deep learning systems are good enough for unmanned spacecraft? How we would generate the learning data would be a question though.
[ Parent
| on HN ]
Re:Computing glitch may have doomed Mars lander
by hvidgaard on Tuesday October 25, 2016 @08:05AM
(#12787157)
You wastely overestimate the capabilities of the equipment we send into space. It is nowhere near powerful enough to power any meaningful deep learning system in an online fashion. Offline has either to high latency, or a problem getting data to seed the initial learning (and even then we have a problem actually running the network).
Re:Computing glitch may have doomed Mars lander
by lorenzhs on Tuesday October 25, 2016 @10:05AM
(#12788076)
Well than and there just isn't a whole lot of training data to play with. It's doubtful that a simulator could ever model actual landing situations well enough that the whole thing will a) work on Mars and b) not overfit weird behaviour in the simulator. Faulty ground sensing has resulted in the loss of more than one mission, and deep learning isn't going to fix any of that.
Oh, and s/wastely/vastly/ ;)