The method starts by defining a concept space. For the demonstration case I chose the problem of a robot recognizing a room from cues.

The second step is to change the concept space into an expert system. I used the Tiny Mycin engine. I verified the expert concluded the correct room from inputs which were on the axes of the concept space.

The next step is to follow a mechanical procedure that I had developed to turn the expert system into a neural network design. To demonstrate the neural network I built it. I used op amps as soma, diodes for the dendritic trees, and resistors for weights. The network functioned as anticipated.

If any op amp in the network were to fail or be unplugged, the equivalent of a rule in the expert would be lost. So I created concept ‘transforms’ that cause each neuron to performing composite work. I furthered this by building an adder where stuck ats would not prevent the adder from failing. It would glitch for a while, improve, and then start producing correct answers again. Such networks required what I called 'praecuro' stimulus, or they would forget the intended function.

This was a joint project for a course in Neural Networks and an advanced course in AI at UT Austin in 1992. I also worked on it further after turning it in. It has affected my thinking about the algorithms used for my startup Reasoning Technology.

It started as a Bayesian network for playing chess in 1988. I noticed it gave results reminiscent of what we saw for solutions to the Schrodinger equation. So in 1993 I turned it around and proposed the Bayesian network as a model for Quantum Mechanics.It wasn't for any class, so the work just remained in my filing cabinet for years. More recently I put it up on Research Gate, The White Knight is Talking Backwards I hope to get more time to work on it more at some point.

My proposed solution was to add a second dimension to the time variable. Time t would tick, then we would step in time u. Consider traveling upstream in a canoe as an analogy, and then coming to a waterfall. We take steps in real world while going up the stream in time t. Between t time steps we take time steps in u. During these time steps we take the canoe out of the water, walk on a smooth trail going around the waterfall, and then set the canoe back in the stream for the next time t tick. In our expanded world there are no sharp corners, every step was over smooth terrain.

This was done for a 1990 course dedicated to writing a Spice Simulator. I asked for and received an extension to work on the simulator over the following summer so that I could get the 2D time algorithm working. It did work, and it brought to light other issues hindering unconditionally convergence, such as the precision of the arithmetic used, and the possibility that multiple choices for moving forward could arise, and potentially a simulator might oscillate between them.

If all of space, including every point, were expanding particles would become ever more crowded. If we scale to that of an observer in this space, the observer will see that particles are inexplicably attracted to each other. I worked on this in 1998, and again I hope to have more time to get back to it. It is also described on my Research Gate page. At one time I had an animation for it on my website. This is and the prior project described on this page, are extensions of my Infinitus project found on the 'youth' page. These are fun things to contemplate.

For one of the electronics design courses we had to build a project. I got a bit carried away, and built a laser scanning instrument for measuring laser scattering on a biological sample for Dr. Motamedi.

The control systems course was all about solving linear differential equations and answer question of stability. It entailed such things as Laplace Transforms and Bode Plots. I wrote a program that solved control systems problems posed in the text book. I made a deal with the professor to turn this in instead of turning in the homework.