On Wed, Oct 20, 2021 at 07:32 PM, Tom, wb6b wrote:
However there are other prebuilt examples that can be downloaded trained and run for practice. If I find anything interesting in my notes from when I discovered this Amazon service existed I can add some links.
It is cool that we have come to the point where there a lot of plug-n-play module that can just be downloaded, and maybe modify a bit with your own training data, and do useful things. Were are entering the Arduino IDE stage of AI/ML development.
Here is a web site that was a very popular read for the folks trying to learn AI/ML in the "assembly language" days of AI.
Probably worth a cursory look at?gradient descent and?back propagation. These are the core of how these systems train. Worth it to try to get a feel for what goes on under the covers, not just the drag and drop pre-made models step we are now entering. Not that the phase we are entering is bad. Its good. I think the snake oil salesmen that tried to promote AI as magic solutions for everything have largely been exposed and gone away?
Also, in the early phase the nonlinear curve that seems to let the "neurons" produce "arbitrary" functions based on the sum of many inputs to a neuron was either an exponential function or based on calculating tangents. Seems very costly for something that is calculated over and over again while a computer preforms a big loop stepping through each "neuron" one by one recalculating the latest value. But the "sigmoid" functions were convenient for the academic early research stage of ML.
Later someone tried a nonlinear function based on a half wave rectifier. Really fast and easy to calculate and seems to work as well as the computationally complex sigmoid functions.
Anyway, I'm prompted to dig back in and start coming back up to speed so I can try some of these new models.
Tom, wb6b
