Sorry this was not how I remember the tweet.
Anyway. I wonder. If you have a stiff problem. One that requires a lot of small time steps. Could this be made into Random Loss Problem. Where use GAN Questions 0/1 somehow and get rid of a lot of loss function checks and updates.
The idea is to use ?residuals (överbliven data) on the layers and a range holder.
To enhance the experience of the raspberry pi 3b+. Use an usb external hdd, not a sd card. Much faster. Tutorial is online (gparted + rsync). Then zram-config via dpkg package. Then lastly Xephyr for that multi desktop experience with XFCE4 for instance.
Maybe this can make airplanes go with much less fuel. Control turbulence with some objects and a magnetic field?
The Most Prominent Recognizable Object In A Small Patch Is A Line. Small Kernels Then Becomes Line Classifiers? A ?real brain like model() could jump back and forth different internal classifiers. Thats my idea.