A new era of computing - just matrix multiplication
It’s been a while since I did write a blog post - professionally and about technology. The question is, why write? Why now?
The true answer is: I don’t know. Just got the urge.
However, the more subtle answer is: to keep up human writing. Or at least, contribute something - albeit a tiny tiny thing.
Another answer could be: to connect the dots.
We are and have ever since been living in changing times. The latest rise of AI reminds me of the late nineties. When the internet got accessible to everyone.
Back then, I did an internship at the university in my home town in an institute for “Neuro informatics”. I helped a PHD student to train a robot arm. The task was to use a Self Organizing Map (SOM or Kohonen network) in order for the robot to learn how to position it’s arms through the three joints so that it reaches a specific position given by x, y, z coordinates. Basically, it was an early form of reinforcement learning with me being the feedback, by leaving the computer, walking into the hallway, measuring the current position of the robot, going back and typing in the numbers to let the SOM do more training iterations. In the end, it was quite precise - quite.
I then stayed with neuronal networks for some time doing some minor projects, like a simple Tic-Tac-Toe player trained via Q-Learning that was, well unsurprisingly unbeatable. I also worked on a routing algorithm for robots based on reinforment learning.
One day during my studies, I had a lesson on systems analysis and theory and the professor that held it, was a seasoned automation engineer by heart. I can still recall his face when he ranted on neuronal networks being just matrix multiplication and a bunch of PID controllers in a row. And how this is well known since the early 20th century. Just do the math!
Well, if I had known that he was totally right, but not as he probably thought.
To be continued…