One of my early "this is neat" programs was a genetic algorithm in Pascal. You entered a bunch of digits and it "evolved" the same sequence of digits. It started out with 10 random numbers. Their fitness (lower was better) was the sum the difference. So if the target was "123456" and the test number was "214365", it had a fitness of 6. It took the top 5, and then mutated a random digit by a random +/- 1. It printed out each row with the full population. and so you could see it scrolling as it converged on the target number.
Looking back, I want to say it was probably the July, 1992 issue of Scientific American that inspired me to write that ( https://www.geos.ed.ac.uk/~mscgis/12-13/s1100074/Holland.pdf ) . And as that was '92, this might have been on a Mac rather than an Apple ][+... it was certainly in Pascal (my first class in C was in August '92) and I had access to both at the time (I don't think it was turbo pascal on a PC as this was a summer thing and I didn't have a IBM PC at home at the time). Alas, I remember more about the specifics of the program than I do about what desk I was sitting at.
Steeeve 1 days ago [-]
I wrote a whole project in pascal around that time. Analyzing two datasets. It was running out of memory the night before it was due, so I decided to have it run twice, once for each dataset.
That's when I learned a very important principal. "When something needs doing quickly, don't force artificial constraints on yourself"
I could have spent three days figuring out how to deal with the memory constraints. But instead I just cut the data in half and gave it two runs. The quick solution was the one that was needed. Kind of an important memory for me that I have thought about quite a bit in the last 30+ years.
aperrien 1 days ago [-]
An Aeon ago in 1984, I wrote a perceptron on the Apple II. It was amazingly slow (20 minutes to complete a recognition pass), but what most impressed me at the time was that it did work. Since that time as a kid I always wondered just how far linear optimization techniques could take us. If I could just tell myself then what I know now...
windsignaling 21 hours ago [-]
I'm surprised no one else has commented that a few of the conceptual comments in this article are a bit odd or just wrong.
> The final accuracy is 90% because 1 of the 10 observations is on the incorrect side of the decision boundary.
Who is using K-means for classification? If you have labels, then a supervised algorithm seems like a more appropriate choice.
> K-means clustering is a recursive algorithm
It is?
> If we know that the distributions are Gaussian, which is very frequently the case in machine learning
It is?
> we can employ a more powerful algorithm: Expectation Maximization (EM)
K-means is already an instance of the EM algorithm.
aardvark179 1 days ago [-]
I thought this was going to be about the programming language, and I was wondering how they managed to implement it on a machine that small.
foobarian 23 hours ago [-]
That's funny, pretty sure we used Standard ML on the old oscilloscope Macs in undergrad. Not Apple 2 of course, but still already pretty dated even at that time (late 90s).
noelwelsh 1 days ago [-]
That's also what I was thinking. ML predates the Apple II by 4 years, so I think there is definitely a chance of getting it running! If targetting the Apple IIGS I think it would be very achievable; you could fit megabytes of RAM in those.
dekhn 1 days ago [-]
Likely any early implementation of ML would have been on a mainframe or minicomputer, not a 6502. A mainframe/minicomputer would have had oodles of storage (both durable and RAM), as well as a compiler for a high level language (which fits what I can see in https://smlfamily.github.io/history/ML2015-talk.pdf and other locations).
noelwelsh 16 hours ago [-]
So I've been mildly nerd sniped. It looks like the first target was a PDP-10 [1]. It ran Stanford Lisp used by the "DEC 10" implementation of ML. The architecture is pretty unusual by modern standards, but it doesn't look to be that powerful and seems to top out at around 1MB of RAM. Next up we have a VAX [2] implementation. It's not clear which specific system it was originally developed for, but we're talking early 80s so it probably wasn't much more powerful than the PDP-10. Either way, I think a maxed Apple IIGS with a hefty 8MB of RAM and perhaps overclocked to 14MHz is more than enough raw power to handle ML. Unfortunately I haven't been sufficiently nerd sniped to actually implement this. I leave that as an exercise for the reader ;-)
an enormous amount of software was developed on the PDP-10 and PDP-11 and later VAX systems that could not have been done on microcomputers in the day. You can't just compare raw RAM and clock rates, the PDPs were set up for multi-user productivity on complex problems and had a wide range of system software to enable building and deploying advanced software.
Scramblejams 1 days ago [-]
Same. What flavor of ML would be the most appropriate for that challenge, do you think?
taolson 1 days ago [-]
While not exactly ML, David Turner's Miranda system is pretty small, and might be feasible:
Applesoft BASIC is just so darn readable. Youngsters have nothing comparable these days to learn the basics of expressing an algorithm without having to know a lot more.
And if it ever became too slow, you could reimplement the slow part in 6502 assembler, which has its own elegance. Great way to learn, glad I came up that way.
drob518 1 days ago [-]
Upvoted purely for nostalgia.
amilios 1 days ago [-]
Bit of a weird choice to draw a decision boundary for a clustering algorithm...
Looking back, I want to say it was probably the July, 1992 issue of Scientific American that inspired me to write that ( https://www.geos.ed.ac.uk/~mscgis/12-13/s1100074/Holland.pdf ) . And as that was '92, this might have been on a Mac rather than an Apple ][+... it was certainly in Pascal (my first class in C was in August '92) and I had access to both at the time (I don't think it was turbo pascal on a PC as this was a summer thing and I didn't have a IBM PC at home at the time). Alas, I remember more about the specifics of the program than I do about what desk I was sitting at.
That's when I learned a very important principal. "When something needs doing quickly, don't force artificial constraints on yourself"
I could have spent three days figuring out how to deal with the memory constraints. But instead I just cut the data in half and gave it two runs. The quick solution was the one that was needed. Kind of an important memory for me that I have thought about quite a bit in the last 30+ years.
> The final accuracy is 90% because 1 of the 10 observations is on the incorrect side of the decision boundary.
Who is using K-means for classification? If you have labels, then a supervised algorithm seems like a more appropriate choice.
> K-means clustering is a recursive algorithm
It is?
> If we know that the distributions are Gaussian, which is very frequently the case in machine learning
It is?
> we can employ a more powerful algorithm: Expectation Maximization (EM)
K-means is already an instance of the EM algorithm.
[1]: https://en.wikipedia.org/wiki/PDP-10
[2]: https://en.wikipedia.org/wiki/VAX
https://codeberg.org/DATurner/miranda
And if it ever became too slow, you could reimplement the slow part in 6502 assembler, which has its own elegance. Great way to learn, glad I came up that way.
[0]: https://proceedings.mlr.press/v170/marx22a/marx22a.pdf
(I mean, the pictures look cool and all.)
IE, did the author want to experiment with older forms of basic; or were they trying to learn more about old computers?
Diffusion, back propagation, attention, to name a few.
Back prop requires and limits to analytically differentiable in a normal way.
Attention is... Oh dear comparing linear regression to attention is comparing a diesel jet engine to a horse.