#garfield Logs

Oct 07 2023

12:24 AM polprog: rue_mohr: you might be right about galvanic isolation
12:37 AM polprog: yes
12:37 AM polprog: but you will have to train it
12:38 AM polprog: the models import the trained network, ie, with weights set on each neuron
12:38 AM rue_mohr: the standard training data is a memory snapshot of an image loader
12:38 AM rue_mohr: polprog, I just reposted something intereting
12:38 AM polprog: you can code your own from the ground up, its gonna be simple but should be doable to make one that can tell a straight line from a curved one or something
12:38 AM polprog: oo, ill see
12:39 AM rue_mohr: fellow says that if you overbuild the network, and initialize it with the right random values, the network you want is in there, you just have to pair it down
12:39 AM rue_mohr: doesn't mean the process of pairing it down is any easier than adjusting the weights
12:39 AM rue_mohr: but
12:39 AM polprog: well, yeah, that is likely statistically
12:40 AM polprog: but then you still have to find the neurons which do the work you want
12:40 AM rue_mohr: I find the idea interesting in the context of working out how the brain would adjust weights of individual neurons
12:40 AM polprog: Im actually starting some neural nework labs this semester
12:40 AM rue_mohr: oh, drag me along would you?
12:40 AM polprog: sure
12:41 AM rue_mohr: skip the crazy math stuff tho
12:41 AM rue_mohr: I'm interesting in the real world realization of it
12:41 AM polprog: not sure how much of it will be coding one from ground up (always wanted to) and how much of it will be "download and run tensorflow"
12:41 AM polprog: the math stuff can be skipped
12:41 AM rue_mohr: yea, that tensor flow thing is a snowball
12:42 AM polprog: a build system written in java, perfect for the winter
12:42 AM polprog: you know how a neural network works right?
12:42 AM rue_mohr: once you start importing the libraries, you also start importing memory snapshots from god-knows-what
12:54 AM rue_mohr: yes
12:54 AM rue_mohr: but I know several operational deviations
12:54 AM polprog: thinking if you could redo your 4 bit relay decoder into a neural network
12:54 AM polprog: you would need 4 input neurons and 7 output ones
12:54 AM rue_mohr: the lambda function seems to be the largest difference in the methods
12:54 AM rue_mohr: haha
12:54 AM polprog: you might need to have 1 7 neuron layer inside
12:54 AM rue_mohr: that relay thing is something I still find amazing
12:54 AM polprog: oh yeah! and its so better than the relay logic
12:54 AM rue_mohr: I'm totally lost as to how he optimized it
12:54 AM rue_mohr: I kinda been trying to do that for years
12:54 AM rue_mohr: not specicially with 7 segment decoding
12:54 AM rue_mohr: I thought adder, but I'v learned that an adder is not what you need to build a computer
12:54 AM polprog: im thinking you could make this network like this, 4 input neurons and the first 7-neuron layer is a 1-of-n encoder, and the output layer is weighted in a way that gives you the pattern given each bit
12:54 AM polprog: Now the problem is that the math becomes necesary. At least matrix/vector math
12:54 AM polprog: To train it out of the system you have to solve a large equation
12:54 AM rue_mohr: there are a lot of methods
12:54 AM polprog: to train it in system, without solving that, hmm..
12:54 AM rue_mohr: back propigation?
12:54 AM polprog: That is something i have to learn yet
12:54 AM polprog: there was something like that, but i forgot how it works
12:54 AM rue_mohr: I'm not 100% on how back propigation is worked out
12:54 AM rue_mohr: but
12:58 AM polprog: its in polish, but readme files should be eay to translate
12:58 AM rue_mohr: ok
12:59 AM polprog: hm, no sources, though
01:00 AM polprog: the net files are interesting, they contain the nn structure
01:00 AM polprog: and its all done by people from my uni, so perhaps i could get the source codes
01:00 AM polprog: anyway, ill copy what is here, maybe it will be useful
01:05 AM polprog: https://polprog.net/rozne1/ircjunk/various/ann_osowski.tgz
01:08 AM rue_mohr: netteach - programy do uczenia i testowania sieci wielowarstwowych - program to test ...
01:08 AM rue_mohr: :)
01:09 AM polprog: yep
01:10 AM rue_mohr: huh, yea no source code
01:11 AM polprog: yeah, i was thinking about other floppy with source code
01:11 AM rue_mohr: so
01:11 AM polprog: (not related to NNs)
01:11 AM rue_mohr: if I take a rom, 2816, and put a counter on it, cd4060
01:11 AM rue_mohr: and have the rom generate serial data for 74595
01:12 AM rue_mohr: the system could output data at about 1.25MHz (parallel from 595's)
01:12 AM rue_mohr: (10MHz system clock)
01:12 AM rue_mohr: I'm going to also tie state machines and NN's togethor at some point
01:13 AM rue_mohr: from a ROM I can get 6 parallel serial channels
01:13 AM rue_mohr: so a total of 48 outputs
01:13 AM polprog: gotta run for a moment
01:13 AM polprog: making breakfast
01:13 AM rue_mohr: has room for 8 inputs, but thats on a 16k rom
01:16 AM rue_mohr: with an extra counter bit and doubling up all the 595, 625kHz output update, and 96 outputs
01:17 AM rue_mohr: output count has always been the problem with my state machines
01:17 AM rue_mohr: the hardware ones anyhow
10:42 AM rue_mohr: oh right, I have a todo list...
10:43 AM rue_mohr: says to change the fish water
12:22 PM aandrew: rue_mohr: I did not know that the cdrom spindle motor had a balancing mechanism
12:40 PM rue_mohr: yep
12:40 PM rue_mohr: you know that rattle when you tip a drive around?
12:40 PM rue_mohr: its the balls rolling around in the balancer
12:40 PM rue_mohr: also adds to the sound when they change speed
12:41 PM rue_mohr: ssss
12:48 PM rue_mohr: dont feel like doing much when I'm cold
12:48 PM rue_mohr: I should go tot he shop where its warm
12:54 PM rue_shop1: its warm here
12:54 PM rue_shop1: now I can do nothing comfortably
01:35 PM rue_shop2: ok, fixed the soldering ircon
01:35 PM rue_shop2: wire broke in a place I'd have not expected
02:40 PM rue_mohr: oh, I just found a 64 pin 9572
02:40 PM rue_mohr: but it looks like I couldn't get it to say hell
02:40 PM rue_mohr: o
02:49 PM rue_mohr: it wont reply
04:51 PM rue_shop2: https://www.amazon.ca/uxcell-Barrel-Hinges-Female-Stainless/dp/B0B18B6JQK/ref=sr_1_29
08:17 PM Tom_L: what are those for?
08:22 PM Tom_L: it was pretty chilly here this morning
08:36 PM rue_shop1: buddy wants to make a custom shower curtin rod
08:39 PM Tom_L: hmm
11:04 PM rue_mohr: so
11:05 PM rue_mohr: http://ruemohr.org/~ircjunk/tempimage/p1290641.jpg
11:05 PM rue_mohr: http://ruemohr.org/~ircjunk/tempimage/p1290645.jpg
11:05 PM rue_mohr: I made a new scale for the meter to measure the RPM
11:08 PM rue_mohr: http://ruemohr.org/~ircjunk/tempimage/p1290642.jpg
11:08 PM rue_mohr: http://ruemohr.org/~ircjunk/tempimage/p1290643.jpg
11:08 PM rue_mohr: http://ruemohr.org/~ircjunk/tempimage/p1290644.jpg
11:09 PM rue_mohr: I stripped some chips of the boards form the phone system
11:09 PM rue_mohr: http://ruemohr.org/~ircjunk/tempimage/p1290632.jpg
11:09 PM rue_mohr: http://ruemohr.org/~ircjunk/tempimage/p1290633.jpg
11:09 PM rue_mohr: I fixed the soldering iron
11:10 PM rue_mohr: and babbled nonsense into twitter
11:20 PM * VanUnamed rue_mohr's chip heist
11:25 PM rue_mohr: :]
11:54 PM aandrew: that reminds me, I was supposed to measure/weigh that box for polprog I think it was
11:54 PM aandrew: have to do that at the end of december when I'm back in canada