AI learns to design

The number of circuits that one can make from, say, thirty 4-pin parts probably exceeds the number of electrons in the universe. A human brain can explore an incredibly large possibility space that no procedural computer program could ever approach.

But LT Spice is great for evaluating and evolving ideas. It complements the more creative but less quantitative brain. I sometimes design by fiddling in Spice, letting it check the reasonableness of my instincts.

I have a couple of recent designs that evolved in LT Spice and work great, that I still don't understand.

On the contrary computers excel at assessing far more designs than we can because they're fast & tireless. We can rule out most options very quickly so we don't need to spend time assessing them. Which is faster? Depends on what's being assessed and how.

Spice sims of random circuits won't design anything useful, but harvest the unused cycles of half the PCs on the internet and it becomes at least possible. Maybe in 50 years we'll all be obsolete.

NT

Conventional computers are horribly slow. They can't even drive a car safely.

Brains are probably quantum computers that can simultaneously superimpose enormousely large possibility sets and collapse the good ones.

There have been attempts to design with computers, or to do the easier task, optimize a given circuit. Not much progress so far.

Using AI to "evolve" designs on real FPGAs finds flaws in the particular devices that it exploits, like building ring oscillators that aren't connected to any power supply net and run on parasitics.

That is to say if you only tell the algorithm to use "minimum number of gates" and don't enforce a rule that they must all be powered in some way it will sometimes find a minimum gate solution regardless that works, on one particular device, for one particular task. "Life finds a way"

Indeed, but neural nets do not work that way, although you can emulate it somehow in a procedural way. In the beginning it is about non-linear weights between elements where data is stored in those weights (variables if you will). The learning process changes those weights but the hardware complexity stays the same while getting ever better 'answers'. (This put right? I did some neural net programming years ago, things have evolved enormous after that).

Our brains only uses a very small subset of its neurons for 'design'.

Yes, but I still like to understand what I am doing :-) OTOH how much in depth _do_ we understand, we do no even know everything about the electron and even less about the universe. So from that realization we will always be tinkering. The steam engine that started the industrial revolution was also the result of just tinkering .. AI can do that without getting tired around the clock...

I like LTspice for filters etc, but there are many good filter design programs on the web.

Also I do think much in blocks that I have ever designed and build and chips I have used. Also more and more is becoming software.

AI designing software???

Now that could create fun solutions.

Indeed, is that good or bad? It may fail if the dies change.. I had those ring oscillators by accident...

If you'd want a design that translates to all devices you'd have to run your genetic algorithm at a higher level of abstraction that only simulates the aspects of the FPGA you want it to be able to leverage.

Algorithms like that I think are still not any good at doing full systems but can definitely be helpful at optimizing sections. The mixed-Chinese postal route problem has applicability to electrical and systems engineering and GAs seem to do better solving it than previous human-devised solutions:

Yes, and neural nets have very different solutions to that sort of problems. A while back a postman was arrested because he never delivered the post, they found thousands of letters and stuff in his house after many people complained about not getting important mail.

Just a few months before that an other postman was caught burying hundreds of letters in the woods. It made the routing a lot easier for them I think.

t's simply not true," said McComb. "

an do more.

ea Spies,

n because they're fast & tireless. We can rule out most options very quickl y so we don't need to spend time assessing them. Which is faster? Depends o n what's being assessed and how.

You jest. An early 70s pocket calculator could do maths faster than we can.

Nor can humans

we don't know

the unused cycles of half the PCs on the internet and it becomes at least p ossible. Maybe in 50 years we'll all be obsolete.

Sure, a lot more compute power is needed. That will be available & cheap on e day.

NT

We have rules that we didn't think to tell the computer. It's a major problem in programming.

NT

But it can't design anything. It can't play table tennis.

Somehow a brain does massive processing and memory and image recognition in fractions of a second, using millisecond logic elements and wet chemistry. It couldn't go that with an array of equivalent logic gates. Nobody has a clue how that works.

I was takling to a neighbor, at our annual block party. He says that he can't remember names or numbers very well, but he never forgets a face that he has once seen, even briefly, decades ago.

Bats echolocate with equivalent ear-to-ear timing resolution in the nanoseconds. Whales and porposes see the shapes of moving objects with their sonar.

Maybe not. Some problems don't get solved much better by applying more and faster Intel CPUs.

Well nobody has ever built a logic array where each element is connected to 500 others.

I've heard claims that the brain uses quantum events to pick options, and the randomness gives the appearance of free will.

But if that's even possible at 98.6 F then why do quantum computers need liquid helium?

I hope you didn't hurt him.

or 10,000 others, which is closer to the fan-out I've heard quoted.

Actual freedom is not necessary anyhow, except to a philosopher. Any behaviour which cannot be distinguished from randomness is sufficient to support the (unnecessary) hypothesis of free will.

Given the exponential rate of error magnification in a chaotic system, even the gravitational pull of a single electron at the far end of the observable universe can be the "butterfly in Brazil", so it hardly helps to even talk about "true" randomness in defence of free will.

Even if the randomness is "true" doesn't make decisions *meaningful"... rather the opposite in fact. The whole philosophical pursuit of quantum randomness as an exercise in defending "meaningfulness" is pointless and counter-productive. It only means something to someone wedded to a Cartesian system where meaning comes across from another realm.

Clifford Heath.

Neural nets were invented in an attempt to explain it. To that extent we do have some clues. Functional magnetic resonance scans of working brains provide a few more. We don't know exactly what the bain is doing, but we can see which areas are doing it.

Cite? Bats can hear up to 80kHz - where the period is 12.5usec.

The speed of sound in air is 343 metres per second, so the wavelength is 4.3 mm.

Nanosecond time resolution translates to a 0.343 micron spatial resolution, and bat labs aren't set up to measure distances anything like that precisely. Neither are bats.

The signals are subject to a mechanical fourier transform and amplitude detection in the ear mechanics before there is even a neural impulse to worry about. Detection of relative phase can then be done at much lower rates, likely by neural correlators on counter-flowing delay lines. All very feasible in wetware without invoking majick.

That is the point, a neural net is principle NOT related to CPUs.

There is hardware developed that behaves much like a neuron.

For beginners:

The thingy in fig 2 can be made from analog components, so no 'puter needed, but 'puters are nice to emulate such thing. The 'weights' storage could be in the form of a capacitor charge, does not have to be digital registers.

Chips exist that contain fig 2 in quantity...

'Learning' changes the weights. Interesting stuff.. There is open source software for Linux to play with this. Those are 'emulations' however, like spice. The analog thing does it have a (CPU) clock?

We are mostly analog, that is why we cannot do big number math easily... Math is just a sub-structure of a few neurons in our brain, and dangerous when taken for reality, much like religion and other 'beliefs' My view anyway :-)

No neural nets are cargo-cult cartoons of a biological nervous system.

Not a bit like a neuron. Even single-cell organisms, with no nervous system, have complex behavior. Things with a couple dozen neurons have extremely complex behavior.

NNs are popular in academia for some reason. I've had job applicants, recent grads, show me their NN projects, which they didn't actually understand.

I do a lot of math at the whiteboard, in my head, which impresses people, but it's analog computing, like a slide rule. 10% or thereabouts accuracy is good enough at a whiteboard, exact results in many cases.

There have been "human computers" who can do high-precision numerical-digits math in their heads.

that's simply not true," said McComb. "

t can do more.

l Tea Spies,

rts

can because they're fast & tireless. We can rule out most options very qui ckly so we don't need to spend time assessing them. Which is faster? Depend s on what's being assessed and how.

an.

your claim was that computers are slow. Even the most primtive CPU IC wasn' t. You've got to go back to the Harwell Witch era to get slow.

mes

my

st the unused cycles of half the PCs on the internet and it becomes at leas t possible. Maybe in 50 years we'll all be obsolete.

one day.

But obviously coming up with lots of random or otherwise circuits & assessi ng them does.

NT