And then there are revisions to the first design to fix environmental effects like temperature and failure rates due to tolerances, adding test points because of the last batch of boards had a failure rate from pick and place errors, etc

"... to perform useful calculations that are within tolerances (close enough)..." I appreciate the sentiment, but a computer that stochastically fails can never be trusted.

I am also curious about why you think ML might be useful in this endeavor. It is well known that current (statistically-based) ML approaches can only generalize within the convex hull of their training, which basically means they can never er come up with anything genuinely novel. In other words, they are not suited to any domain requiring innovation.

@𝕕𝕚𝕒𝕟𝕖𝕒 🏳️‍⚧️🦋 That's where the ML comes in. It can figure all that out by specifying a temperature range that results are considered good. There's nothing stopping us from training the model in a range of temps/humidity.

a computer that stochastically fails can never be trusted.

We don't need to trust it. We need it to "do good enough to get the job done," as is the case in biological systems.

^ Sorry missed the @John Hummel

No worries, @Richard. I committed the same error 😀

"We don’t need to trust it. We need it to 'do good enough to get the job done,' as is the case in biological systems."

I respectfully disagree. For intelligence of the "do good enough to get the job done" variety, we already have plenty of biological intelligence. All the way back in the 1600s, Leibniz realized that the reason to use machines to execute computable functions is precisely that they can be relied upon to do so perfectly. And this is exactly how we use them today. If you ask your computer to tell you the square root of 2, you trust its answer. By contrast, you wouldn't be wise to trust the answer a random person on the street would give you. (The less important reason we rely on computers is that relegating the work to a machine takes the burden off people.)

If I believed my computer would only perform arithmetic operations correctly 90%, or even 99% of the time, then I would never trust its results: I wouldn't trust it with payroll, or scientific computing, or much less any kind of engineering where lives are at stake.

Computers need to be as close as possible (Footnote) to perfect at what they do--because that's precisely why we use them in the first place.

Footnote: Of course no machine can ever be 100% perfect. There's a famous case of a computer in the UK making an error counting votes because a gamma ray just happened to bombard one of the machine's bits, flipping it from a zero to a one and thereby erroneously adding a large power of two to the correct value. But gamma ray strikes are exceedingly unlikely (especially within the bounds of the Earth's atmosphere) and virtually unpreventable. Under no circumstances should we add preventable errors to our computing devices. Especially when you consider just how many calculations go into any mundane task (often millions, billions, or more), even a small likelihood of an individual error can become a near certainty of an incorrect final result.

That last comment made no sense. Let me try again. Imagine you're a leader over an army of engineers and human calculators in Soviet Russia. All you have are parts approved for production by the CCCP - say, knock-off clones of 2N2222's etc. Replace the engineers and human calculators with ML. Replace the CCCP with neoliberal American style markets. Replace the 2N2222's etc. with whatever is selling for cheap these days.

@John Hummel wrote:

I respectfully disagree. For intelligence of the “do good enough to get the job done” variety, we already have plenty of biological intelligence.

But I don't have silicon based artificial intelligence (with the environmental advantages that silicon provides over carbon) guided by something much faster than organic evolution.

I'm not wealthy, powerful, or elite. I'm a knucklehead with weird dreams. I don't have access to DNA synthesis, molecular printers, or any of that.

Thanks to neoliberal American style markets I do have access to mouser, digikey, etc. and whatever tech happens to get wealthy shareholders the greatest ROI.

Thanks to OSS I have access to ML.

So this is very much a peanut butter and chocolate exercise.

Perhaps it speaks of my fatigue (it is now about 1:30 am in Rotterdam, where I currently reside), but your most recent comment made even less sense to me than the one before it. Are you saying, "Imagine a world where ML replaces engineers and mathematicians, and where neoliberals and other fascists reign" rule?

If that's what you're saying, then my answer is simple (and perhaps what you intended): I sure a hell hope the fascists use ML for all their design, engineering, math, reasoning, etc., because if they do, ten they will verily fucked and we will defeat them easily.

It occurs to me that three concepts that are perhaps easily confused have played important roles in our exchange: (1) biological intelligence, (2) computers/algorithms, and (3) ML, which happens to be instantiated in versions of (2).

(1) Biological intelligence evolved in response to the pressures of survival and reproduction: Inferences must be fast (i.e., based on very few training examples) and readily generalizable. Biological intelligence cannot afford to obey the statistical Law of Large Numbers.

(2) Algorithmic/computational intelligence is concerned with (2a) what can be known with absolute certainty (i.e., what is computable), and (2b) how efficiently (in terms of runtime) can we know it?

(3) Modern approaches to machine learning are all statistical, meaning that (unlike biological intelligence) they cannot violate the Law of Large Numbers. For this and other, more technical reasons, they (in contrast to biological intelligence) require enormous amounts of training before they can do anything intelligent. And for technical reasons understood by probably fewer than 100 people -- certainly fewer that 1,000 -- these approaches are fundamentally incapable of generalizing as flexibly as biological intelligence.

One of the advantages of statistical learning over biological learning is that the former can tell you exactly how much to trust your conclusions. One of the advantages of biological learning over statistical learning (also one of its disadvantages) is that it doesn't really give a shit about that. And one of the advantages of human learning over most forms of biological learning is that its capacity for generalization is orders of magnitude greater than that of other biological intelligences.

But this advantage of human intelligence is not only ignored, but actively eschewed, by statistical learning.

People in ML are coming around to realizing how important this capacity of human learning is. But they are so wedded to their approaches that they are incapable of seeing the solution--a solution that requires them to abandon almost al their assumptions.

If you knew me, you would trust me when I say: Modern ML is not the solution to anything.

Based on your last comment, @Richard, I am coming to fear that I have completely missed your original point. If so, I apologize for all the drivel.

I guess the point is that I want to do things just because.

The reality is that I am a United States Wage Slave. I am an inmate incarcerated within the razor wire fences of USian society. Like any inmate, I have to make do with whatever I can get my hands on and utilize in the most creative ways I can manage.

@John Hummel It's okay. I'm used to smarter, more knowledgeable people going "WTF" to my attempts at communicating what's in my head. Most of the time I'm trying to explain the obvious (to domain experts) in an ignorant way or I'm just flat out wrong.

There's nothing wrong with this idea. Plenty of analog circuits are used for things where they are good enough. For example, an analog circuit is used to control intermittent wipers. Sure, the rate is affected by temperature and such, but ... who cares? It's only a problem if you are doing a TV show or something where you want to have multiple cars swing their wipers in sync.

@Isaac Kuo I’m interested in the weird circuits that use parts in weird ways.

I learned about common collector, common emitter, common base circuits as a kid thanks to my Tandy Science Fair 200 in 1 Electronic Project Kit.

But there are all kinds of weird non-linear behaviors in the various curves that can be utilized. This is where ML comes in handy. Most manufacturers are aiming for consistency in the linear regions. Thus ML can speed up convergence on a real-world solution, especially in transistor networks.

Also there's circuit impulse response. Shoot training signals into various parts of the circuit, see what comes out at other various parts and see if you can converge on a solution from that.

I'm mulling over is what minimum "cell" building block circuits should be. In the boolean domain the NAND gate is king. What is the equivalent for the analog domain?

In FPGA the popular one is a LUT, but there's other blocks like whole ALU, DSP, and even CPUs in the bigger ones. For radar which I have experience with, all analog is converted to digital and math blocks do the analog in the digital world. The speed of the ADCs and DACs can be crazy fast in the gigahertz range. That was several years ago, I'm sure it's advanced even further today...

@𝕕𝕚𝕒𝕟𝕖𝕒 🏳️‍⚧️🦋 Yup. But FPGAs cost $5-20 per. In my dream the tech was secret 1970's skunkworks level elite. I was asked if I could repair a robot with an inflatable bipedal shaped top and a bottom made of lots of interlocking metal parts attached to a cloth like substrate with solenoids on xyz axes. Each metal piece had a PCB with a few transistors, capacitors and resistors that did not appear to be the usual circuits. The bottom half appeared to be capable of scurrying around like a millipede only with a matrix of feet, manipulating nuts and bolts, grasping tools, etc.

The implication was that some engineering team and their human computers figured out the absolute minimum discreet part count necessary to automate the possible movements and temperature+pressure feedback signals that each interlocking metal part actuated by the solenoids require to accomplish the mission.