#robotics Logs

Oct 06 2018

12:05 AM zhanx: if you have too many layers its slow to learn also
12:09 AM rue_mohr: right
12:09 AM rue_mohr: but one of the neat things
12:09 AM zhanx: ok ordered more filament
12:09 AM Jak_o_Shadows: hmm. What are the imporant wires in an OBD connecotr?
12:09 AM rue_mohr: is that the start weights can be random, and it works it out
12:09 AM rue_mohr: and that static value can be somewhat random too
12:09 AM zhanx: Jak_o_Shadows, power, don't cross it to ground a lot of the old ecm's failed to protect for that
12:09 AM rue_mohr: but, how do you train a multilayer...
12:09 AM rue_mohr: this excel example has no hidden layers
12:14 AM Jak_o_Shadows: yeah, 9
12:14 AM Jak_o_Shadows: So i'll ditch one of them, and just use some ethernet wire
12:16 AM zhanx: my montero only uses 7 so go figure
12:16 AM rue_mohr: it cant learn exclusive or
12:18 AM rue_mohr: the static input value acts like a bias
12:18 AM rue_mohr: without muffing up the math
12:28 AM * Jak_o_Shadows hopes those reed switches were normaly closed
12:32 AM Jak_o_Shadows: oh, i got dual position reed switches
01:25 AM rue_mohr: hwo can we generate an optimized system for storing neural weights
01:26 AM rue_mohr: wrong target
01:26 AM rue_mohr: we need to know what neurons were storing the weights between
01:26 AM rue_mohr: what about that voxel tree whats-his-name was using
01:27 AM rue_mohr: you think you can make that express a neural network link map?
01:27 AM rue_mohr: maybe, that or soemthing like it
01:28 AM rue_mohr: so, the origional neural networks I made were close, but lacked the threshold functionality
01:28 AM rue_mohr: I think you also made them as a feedback mesh
01:28 AM rue_mohr: I was on to soemthing :)
01:28 AM rue_mohr: your wter is boiling...
01:28 AM rue_mohr: !!
01:40 AM TheSolenya: hey, does anyone have resources on machine learning for robotics?
01:40 AM TheSolenya: I'm having trouble finding any papers on the subject that aren't paywalled
01:50 AM rue_mohr: did you say neural netwoks?
01:50 AM TheSolenya: no, but that's what I meant:P
01:51 AM TheSolenya: ML is a science field of NN, but NN is not necissarily ML
01:51 AM Jak_o_Shadows: I would've put it the other way around myself
01:51 AM TheSolenya: Well, you can make a neural network that doesn't use any ML theory
01:52 AM Jak_o_Shadows: But like, machine learning could basically just be simple curve fitting
01:52 AM TheSolenya: but ML is all about using neural networks in specific ways, like building, training, and running
01:53 AM TheSolenya: yeah... it could, but in this case I'm trying to program a robotic arm with a claw, and 3 cameras to track the area of interest for the robot
01:53 AM rue_mohr: soyou want something better than tf
01:53 AM TheSolenya: Well, I'm using pytorch right now, which I feel is better than TF, is there something better than that for this type of task?
01:54 AM rue_mohr: well, asembler
01:54 AM rue_mohr: but whatever
01:54 AM rue_mohr: what task are you taking on?
01:54 AM TheSolenya: asembler, as it, write assembly language? or is that a suite
01:54 AM Jak_o_Shadows: you could call me a machine learnign skeptic, ther'es too many buzz words for me to be comfortbale
01:55 AM TheSolenya: rue_mohr, Just learning right now.... hoping in a decade or 2 to have a sentient robot with general intelligence
01:55 AM TheSolenya: for now, I just need to learn basics, then I can do more specific stuff that's more advanced
01:55 AM mrdata_: you'll have to solve the catastrophic forgetting problem
01:55 AM mrdata_: for AGI
01:56 AM rue_mohr: well, you ahve to have a target to make a ml system
01:56 AM TheSolenya: yeah... I have an idea for how to do long-term memory using NLP, but for now I'm not trying to do AGI, that's just my end goal, for now, i want to learn basic robotics related ML techniques, then i will step it up once I get it
02:00 AM TheSolenya: I have a robot, if that's what you mean by target, and my goal right now, is to learn how to make it stack toy blocks
02:00 AM rue_mohr: hmm
02:00 AM rue_mohr: stm32
02:00 AM Jak_o_Shadows: ah, damn, I forgot the bleeder resistor
02:00 AM rue_mohr: what do you want to make your robot do?
02:00 AM TheSolenya: I want to make it play with things
02:00 AM TheSolenya: pretty much, just be a stupid robot that plays with toys all day
02:00 AM rue_mohr: whats the target that generates positive feedback to the network
02:00 AM TheSolenya: cameras for input(got 3 hd cams), and successfully manipulating objects in relation to other objects within it's reach, would be a low-loss result
02:00 AM TheSolenya: Basically I just want to learn the most simple ML for robotics stuff, like hello world for ML motor control
02:00 AM TheSolenya: and then I can ask for more advice once I do that
02:00 AM rue_mohr: you know rue, you can have a trinary system that generates positive, null, or negitive feedback
02:00 AM rue_mohr: just tell me how to train hidden layers already
02:00 AM TheSolenya: I'm a noob to both robotics, and to ML, but I have a robot and I have time to learn what I need to learn, and I want the skills I gain from this experience to benefit my end-goal of AGI
02:00 AM TheSolenya: rue_mohr, in what API?
02:01 AM rue_mohr: generic
02:01 AM TheSolenya: I know how to build ML models, but I'm not experienced enough to be trying to help someone with a question like you just asked
02:01 AM rue_mohr: https://www.youtube.com/watch?v=3993kRqejHc
02:01 AM rue_mohr: ^^ no hidden layers
02:01 AM TheSolenya: That's amazing!
02:01 AM mrdata_: https://arxiv.org/abs/1805.07441
02:01 AM rue_mohr: tho, there are only two patterns that could be derived for it using hidden layers
02:02 AM TheSolenya: you mean you don't need ML to use excel?
02:02 AM mrdata_: Overcoming catastrophic forgetting problem by weight consolidation and long-term memory
02:02 AM TheSolenya: is weight consolidation the new law they found where you balance weight sums to avoid cascade overload failure?
02:03 AM TheSolenya: I have a theory on how to do long-term memory, but I've not tried to implement it yet, so it's still just a theory
02:03 AM TheSolenya: It's too complex to explain here, I'd spam you with 3 or 4 pages of text
02:05 AM TheSolenya: I'm not ready to try and implement my long-term memory idea, I have too much more to learn that I need to implement before I can get there
02:06 AM TheSolenya: but I feel like it's a pretty good approach, uses NLP in funky ways that I've not seen people use NLP for
02:06 AM TheSolenya: the question is... am I smarter than I think, or am I bound to fail and try something else like I expect to
02:07 AM rue_mohr: I say try it
02:07 AM TheSolenya: rue_mohr, I intend to, once I'm done learning the building blocks I need to implement it
02:07 AM TheSolenya: The only way to know if it's a workable idea is to try it, but I need a foundation to add it to
02:08 AM rue_mohr: :) excel
02:08 AM rue_mohr: or in my case, gnumeric
02:08 AM TheSolenya: lol... I don't know excel, and I highly doubt excel can handle billions of trainable parameters
02:09 AM rue_mohr: funyn , an hour ago I thought of a way to collapse hidden layers
02:09 AM TheSolenya: collapse them for what purpose?
02:09 AM rue_mohr: memory and compulation speed
02:09 AM TheSolenya: oh, so like, extracting algorithms from a trained model?
02:10 AM rue_mohr: as many hidden layers as you want only take up the memory of 1
02:10 AM TheSolenya: I read a paper on doing that a couple weeks ago... but I didn't bookmark it and I forgot
02:10 AM TheSolenya: Oh, that's cool
02:10 AM rue_mohr: but now I need to learn how to train hidden layers
02:10 AM TheSolenya: sounds too out-there to be feasable but I've been wrong before!
02:10 AM TheSolenya: you should go through the pytorch documentations tutorials
02:11 AM rue_mohr: na
02:11 AM TheSolenya: they are very good docs and you don't need to be an ML expert to do them, and your questions will be answered
02:11 AM rue_mohr: so far everything sucks,
02:11 AM rue_mohr: why the hell should I need a virtual machine to play with a tf network?
02:11 AM TheSolenya: yeah, everything does suck, but you gotta learn an API if you want to test your theories... i recommend pytorch
02:11 AM rue_mohr: like, seriosly people
02:11 AM TheSolenya: virtual machine?
02:12 AM TheSolenya: you got PCIe passthrough on KVM or something?
02:12 AM rue_mohr: yea, what the hell eh?
02:12 AM rue_mohr: anyhow, if I can work out how to teach a network, I'm good, everything else is s sinch
02:12 AM TheSolenya: you should have a bare metal linux install with a very slim setup to support your ML API and your GPU
02:12 AM rue_mohr: no
02:13 AM TheSolenya: why not?
02:13 AM rue_mohr: my target is an stm32, @ 72Mhz
02:13 AM TheSolenya: ahh
02:13 AM rue_mohr: the technology exists
02:13 AM TheSolenya: My target is 2 GV100 Quaddro's
02:13 AM TheSolenya: so like 10,000 cores and 64gb of ddr6
02:14 AM rue_mohr: I'm working on this to port myself onto an electromechanical platform, but thats for later
02:14 AM TheSolenya: port yourself?
02:14 AM TheSolenya: I don't get it..
02:14 AM rue_mohr: yes, the most ugly software port ever
02:14 AM rue_mohr: and my existance depends on it
02:15 AM TheSolenya: I don't get it, you want to port your brain's algorithms and data into software?
02:15 AM rue_mohr: the gpu's are powerfull enough, but dont have enough memory yet
02:16 AM TheSolenya: yeah, GPUs need more RAM, the GV100's have 32gb... I've got 1 and am waiting for some other hardware to sell so I can buy another
02:16 AM rue_mohr: unless I can work out a amazing way to express the network
02:16 AM TheSolenya: I'm trying to find someone to buy a really expensive computer, but most people don't want to pay $5000 for a crazy computer when $1000 will run anything average people will ever touch
02:17 AM TheSolenya: I paid over 10k for that PC, but used, and 7-8 months old/outdated, I hope to get 5k for it
02:17 AM rue_mohr: ok back on the search for a down-to-earth howto on training hidden layers
02:17 AM TheSolenya: I need another GV100 more than I need a big beast for compiling stuff
02:18 AM TheSolenya: so, the 1950x is up for sale in all it's glory... When I built this computer I named it Tremendous Overkill
02:19 AM TheSolenya: I'm taking a 50% loss for that 7-8 months of use, at my asking price... but I want to be able to train 2 different ML models at the same time, because I have one project, and am about to start with the second project which is my robot
02:20 AM rue_mohr: and your going to let me beat you witha $3 microcontroller?
02:20 AM TheSolenya: lol, sure
02:20 AM TheSolenya: this is really a learning and research project
02:21 AM TheSolenya: Like, I started with ML about 7 months ago, and I just ordered the final part for my robot... got the servos in the main arm working with a rpi using gpio pins
02:21 AM TheSolenya: just programmed it with python using the pi's library from their site
02:22 AM rue_mohr: stm32 and 4x 4017 (total about $8) can run 24 servos
02:22 AM TheSolenya: I think i went too cheap on the robot though, I only spent like $500 in total for the arm, the hand, and the cameras
02:22 AM TheSolenya: cool
02:22 AM TheSolenya: but I want to learn to do stuff with ML, not just control robots using hard-coded functions and variables
02:23 AM TheSolenya: I can already program a robot with a dev board and some C or python, I've done that long time ago... Now I want to create something that can be integrated into other projects once I build it and learn how to make it do what I want
02:24 AM TheSolenya: Like, the thing I'm asking for links on how to do, is not to just control a robot, it's to use ML to control a robot in a way that it can learn and improve itself... obviously that statement does not reflect what I will first code to learn, but that's a goal
02:25 AM TheSolenya: I have found a lot of papers on it, but ALL are behind a paywall
02:25 AM TheSolenya: So I just see the abstracts and get teased by them
02:27 AM rue_mohr: you know what an fft is?
02:27 AM TheSolenya: um, sort of.. like a robotic assembly line right?
02:28 AM rue_mohr: no
02:28 AM TheSolenya: oh, then no
02:28 AM rue_mohr: fast fourier transform
02:28 AM rue_mohr: hmm
02:28 AM TheSolenya: oh, okay, yeah I know fourier transforms sort of... I mean I used to know them, and vaguely remember some stuff
02:28 AM TheSolenya: I haven't studied them since I learned them 8 or 10 years ago
02:28 AM TheSolenya: and I haven't used them
02:29 AM rue_mohr: well, I implemented an fft in 8 bit integer on an 8 bit avr
02:29 AM TheSolenya: so I could refresh quickly and know them again, but I couldn't use them without at least a few hours of review
02:29 AM TheSolenya: that sounds cool
02:29 AM rue_mohr: which I had lots ofpeople telling me coulnd't be done
02:29 AM rue_mohr: I can make a pretty complex nn on an stm32
02:29 AM rue_mohr: $3 processor
02:30 AM TheSolenya: people told me I couldn't make a mobile linux platform because there was no public hardware documentation for any of the phones I was targeting... I have everything working now except GSM/LTE radio, and the wayland port is pretty buggy
02:30 AM TheSolenya: they said I wouldn't get hardware acceleration on video or rendering, and it took me < 3 months to get it working
02:31 AM rue_mohr: nice
02:31 AM TheSolenya: yeah... I've worked over 3 years to get linux 4.15 running on 7 different phones, and I have a tool I made to help automate the porting of the kernel to new devices using their android kernels
02:31 AM TheSolenya: I had a employee helping for 2 of those 3 years
02:32 AM TheSolenya: the Ubports (Ubuntu Touch) people called me crazy, said I was being stupid trying what I was doing... so they are gonna be left behind in a few years when nixphone rolls out
02:33 AM TheSolenya: it's way ahead of my projected ETA, i figured 10 years to stable 1.0 release, but it's been 3 and I'm now thinking 2 more for everything except the compiler, which is not necissary because you can run elf binaries compiled with gcc
02:33 AM rue_mohr: can I run it on old blackbery smartphones?
02:34 AM rue_mohr: http://neuralnetworksanddeeplearning.com/chap5.html <-- I'm reading...
02:34 AM TheSolenya: No, right now most of my development is on the Nexus 5, but I have a few other nexus phones, and a few samsung phones, that I'm targeting
02:34 AM TheSolenya: Eventually I want my platform to ship with devices, but for my 1.0 stable release my goal is just to support the 7 phones I've been developing on as alternative non-android ROMs
02:50 AM rue_mohr: huh
02:50 AM rue_mohr: I think I know how to deal with that
02:51 AM TheSolenya: with what?
02:51 AM TheSolenya: the annoying guy who just joined the channel? (me)
02:52 AM rue_mohr: with an overall factor that tries to equalize all the weights
02:53 AM rue_mohr: I'm gonna get some slep
02:54 AM TheSolenya: later man! sounds like a plan
02:56 AM flav0r: anyone here ever looked at the inside of an old game controller?
02:56 AM flav0r: i just opened up an old nintendo controller to get the dpad from it
02:57 AM flav0r: and where it contacts theres some sorta coating on the copper
02:57 AM flav0r: looks like some sort of carbon buildup
02:57 AM flav0r: probably doesnt corrode
03:11 AM Jak_o_Shadows: and all the electrical is done for this reversing camera
03:11 AM Jak_o_Shadows: now just gotta mount it
03:11 AM Jak_o_Shadows: and by "reversing camera", I mean the screen for one
08:50 AM rue_mohr: its interesting really,
08:50 AM rue_mohr: the neural networks are amazingly stable
06:02 PM flav0r: show me your fake brain dude
11:12 PM durrf is now known as ziggggggggggy