Yolol chip neural network possibility

Fox

Endokid
Joined
Mar 12, 2020
Messages
3
#1
We all know That yolol codes are executed 0.2s per line, executing a complex program would take forever. and in most cases, the latency would make the program meaningless. How ever, what if we build a neural network with individual yolol chips function as a neuron? Would that make complicated programs possible?
I don't know much about deep learning but as far as I know, in a neural network, a neuron only calculates the activation function of ((bias) plus the sum of ((value of neuron in the previous layer) times ( weight of the neuron in the previous layer))).
Can we construct a relay of yolol chips with layers and make them run simple addition and multiplications simultaneously? Each neuron would take only one line to calculate its function, and a 5 layer network will only take 1 second to complete a task that would take minutes even hours to run if written in a traditional form.
As for the training of the neural network, we can do it somewhere else outside the game and enters the bias and weights by hand into the memory chips.
It will not be feasible for small ships or maybe only works for big stations running by factions with loads of work force, but is it really possible in-game?
 

Burnside

Master endo
Joined
Aug 23, 2019
Messages
308
#2
each sequence is still going to be reliant on the latency of the last set of weights updating, you'd be better off looking at chipless or near-chipless solutions, abusing the update speed of device fields and state switching to create some form of logic gate system, but that'd be a physical computer with buttons or sets of buttons as gate components- it'd be huge
 

Fox

Endokid
Joined
Mar 12, 2020
Messages
3
#3
each sequence is still going to be reliant on the latency of the last set of weights updating, you'd be better off looking at chipless or near-chipless solutions, abusing the update speed of device fields and state switching to create some form of logic gate system, but that'd be a physical computer with buttons or sets of buttons as gate components- it'd be huge
It would be better for simple calculations with low-latency requirements but would be too complicated for applications that require numerical calculations, like image recognition, radar systems, auto-piloting, Inverse Kinematic solver for industrial arms etc.
 

Burnside

Master endo
Joined
Aug 23, 2019
Messages
308
#4
something like image recog is likely outside of the realm of feasibility for this game, you'd need a visual sensor to start and then design and code an optical intelligence and fit it within yolol/device-field-computation constraints

crude forms of range-limited radar are already plausible with rangefinder arrays, the options and flexibility only increase once the devs release more types of sensors

autopilots are possible in some formats, mostly autobuses on station positioning networks, landing and terrain-following pilot-assist programs are also plausible- things like a combat-oriented autopilot are probably out of the question tho
 
Joined
May 12, 2020
Messages
2
#5
Training the neural network outside of the game
Then recreating it in the game using one chip per one neuron
Assuming you would find a way to get each layer to run slightly after previous one (like enable the chip 0.05 seconds later)
You could actually get it to work at 5 Hz frequency

Unless all chips in game update at the same time, I mean it's based globally, not since the chip was enabled, then the idea MIGHT still work but it would create issue of let's say 1 layer would take 0.2 sec so Your outputs in 5 layer network would be calculated 1 sec after getting an input that might have changed already and also assuming putting constantly new inputs every 0.2 sec it and feeding it forward the network it could cause sync issues

But it's still interesing idea and I will surely take a look into it
 

Burnside

Master endo
Joined
Aug 23, 2019
Messages
308
#6
It could work well in a factory setting, at the very least, but somehow it feels like this network would be massive.
 

XenoCow

Master endo
Joined
Dec 10, 2019
Messages
562
#7
It could work well in a factory setting, at the very least, but somehow it feels like this network would be massive.
That is one problem with some of the more complex programs that take many chips, "physical" space. You don't want half your ship taken up by YOLOL chips, it would not leave room for stuff to act on the envionment, engines, generators, weapons, etc..
 
Joined
May 14, 2020
Messages
4
#8
It would be better for simple calculations with low-latency requirements but would be too complicated for applications that require numerical calculations, like image recognition, radar systems, auto-piloting, Inverse Kinematic solver for industrial arms etc.
It also sounds horribly ineffecient, trying to use the in-game coding from a game to build a network like this. You're essentially adding a whole new dimension to the design of your neural network and countless factors and as you mentioned before, line exec's only happen every 0.2 seconds. A Youtuber I watched recently had made a self driving car in GTA that used a Titan X to run off and this card had almost 100% utilization. It would be interesting to use YOLOL for basic collision avoidance and self driving miners etc. but actual neural networks do not seem possible sadly.
 

Fox

Endokid
Joined
Mar 12, 2020
Messages
3
#9
It also sounds horribly ineffecient, trying to use the in-game coding from a game to build a network like this. You're essentially adding a whole new dimension to the design of your neural network and countless factors and as you mentioned before, line exec's only happen every 0.2 seconds. A Youtuber I watched recently had made a self driving car in GTA that used a Titan X to run off and this card had almost 100% utilization. It would be interesting to use YOLOL for basic collision avoidance and self driving miners etc. but actual neural networks do not seem possible sadly.
I won't expect the network to run autopilot, and that thing would probably only be able to run on a dedicated ship or station that can host such a huge network. We can probably run a little network with around five layers at most and it probably should be sufficient for simple tasks like ship recognition with sensor array or other stuff.
 
Joined
May 14, 2020
Messages
4
#10
I won't expect the network to run autopilot, and that thing would probably only be able to run on a dedicated ship or station that can host such a huge network. We can probably run a little network with around five layers at most and it probably should be sufficient for simple tasks like ship recognition with sensor array or other stuff.
Sounds like a plan to me I might be able to help out with that as soon as I get access :)
 
Joined
May 19, 2020
Messages
12
#11
I think the term he is looking for is "parallel processing" not "neural network". Parallel processing would very much be a thing for any complex task, such as target position prediction on turrets or crude radars.
 

MrFaul

Well-known endo
Joined
May 20, 2020
Messages
82
#12
Yeah a NN won't happen in this game, you would be fighting countless race conditions before you could get anything done.
I also imagine you would need a lot of resources just to get the parts.

Besides NNs aren't a cure all, reliable good engineered formulas will get you oftentimes faster to the goal.
Best example Tesseract OCR.

@Fox just out of curiosity have you ever written a neural net from scratch?
For those interested in the process look here.
 

Mr.Silver

Well-known endo
Joined
May 16, 2020
Messages
58
#13
Yeah a NN won't happen in this game, you would be fighting countless race conditions before you could get anything done.
I also imagine you would need a lot of resources just to get the parts.

Besides NNs aren't a cure all, reliable good engineered formulas will get you oftentimes faster to the goal.
Best example Tesseract OCR.

@Fox just out of curiosity have you ever written a neural net from scratch?
For those interested in the process look here.
Never know what the future holds
 

riley

Learned-to-sprint endo
Joined
Jan 23, 2020
Messages
24
#14
Yeah a NN won't happen in this game, you would be fighting countless race conditions before you could get anything done.
I also imagine you would need a lot of resources just to get the parts.

Besides NNs aren't a cure all, reliable good engineered formulas will get you oftentimes faster to the goal.
Best example Tesseract OCR.

@Fox just out of curiosity have you ever written a neural net from scratch?
For those interested in the process look here.
I agree with the majority of this although i think it may be possible.

Train the network outside the game and hardcode the weights into the chips itself, don't bother with image recognition (too many calculations compared to simpler networks) and don't rely on high speed at all and it should work. Keep in mind YOLOL uses 64 bit fixed point decimals, meaning that accuracy will be limited to those three decimals. I don't see a practical reason to make this though other than saying "Look I made a NN in game".
 
Joined
Jul 10, 2020
Messages
1
#15
I can see using one to play the market eventually, especially if you can set up an autobuy/autosell in YOLOL. That's not a situation where seconds will be as crucial, and especially if you could pull market data for training, train the net outside of the game, and then hardcode the results into the YOLOL, I can see using one. Yes, it might take thirty or three hundred seconds to process a situation, and, say, set a price for an amount of a resource, but that won't be an enormous cost like it would be on a turret or something. Plus, a market is a situation where it won't cost you to have a large, static, multichip machine.
 
Joined
Jul 18, 2020
Messages
6
#16
I highly doubt this will ever become a reality due to the limitations of the language itself and that frozenbyte might ban them due to them toasting the servers
 
Joined
Aug 26, 2020
Messages
2
#17
It seems not feasibile since even with paralel programming it should be a literal hell to optimize networking side. If the code is executes in server side, it might create immense weight on cloud computation also it allows further player abuses. Probably one of the best way to run a proper neural network is sustain additional API for the game so we can use in-game data for training neural networks outside the game and compute such programs locally instead of cloud computing. I am aware that It might create hack oppurtunities but it can be over come by basicly runing essential code in cloud meanwhile commands are given by locally ran programs.
 

Venombrew

Master endo
Joined
Aug 9, 2019
Messages
369
#20
lol some of this stuff is pretty complicated, good thing my only job is to figure it out how to use it after i steal it! i let you brains figure out how to get it to work.
 
Top