Comments by "Immudzen" (@Immudzen) on "Future Computers Will Be Radically Different (Analog Computing)" video.

While the video is largely accurate for classification type models it is not accurate for regression model. I work on neural networks to approximate a PDE solver and a mean square error of 1e-7 is too high to be able to use the network and 1e-8 is borderline. The network I use now has an error around 1e-9. These networks are not sexy, they don't get a lot of nice videos because they don't generate nice pictures usually. However, they are critically important. Calculations that would normally take years can be done in days. NVidia and AMD even build special neural network GPUs for calculating these kinds of networks at high precision. All of my calculations are done in double precision and even changing to single precision significantly degrades the network.
1
Most scientists do think about this. The issues are complicated because if you write a paper on what you learned anyone can take that paper, without your help, and use it for good or bad things. If you don't write the paper you either don't graduate or you lose your job as a researcher. I have been working on neural network to approximate complex simulators. It is designed to make medicine easier and some calculations that would have taken years can be done in days. I am sure that someone could find a way to abuse it. I don't know what that way would be but when it comes to being destructive humans are pretty creative.
1
@teguh.hofstee You end up with a dying neuron problem with that. On deep networks there is a reason that activation functions like swish have become more common.
1