We know the laws of physicals well enough to make large, complex machines. As Engineers, sometimes compromises need to be made to factor in costs, labor, manufacturing constraints, logistics, etc. Good Engineers make reasoned decisions for the way they build things and then run simulations to test reliability. Can't this process be better done with Deep Learning? It could generate multiple solutions for an Engineer to choose from and even have the budgeting worked out in advance. It could then be run through custom simulations before it runs another iteration until it generates the ideal solution.
Machine Learning in Engineering (Construction, Machinery, Aerodynamics)
Kayden Robinson
Other urls found in this thread:
Joshua Powell
install gentoo
Xavier Bennett
*Physics
Xavier Torres
Oh wow, OP. You're the first person to ever think of this!
I'm sure you've thought about how to train this deep learning system you thought up.
Jackson Perez
It doesn't necessarily need to be a complete solution but there are many narrow domains where it could shine. Let's say when you're designing something like a 4-axis robot, the CAD software could automatically include an adaptive stability design mod that a machine learning algorithm could come up with multiple options that you can quickly select and then continue working on it.
A ton of money is spent on Natural Language Processing and Speech Recognition and Synthesis to make shitty botnets and surveillance. A CAD is more limited in scope and is a precisely defined environment. Why aren't CAD development companies including any features? It doesn't have t be world-changing but surely there are some areas that can be improved on? CATIA already charges you $60000 for a license. What about Electronics Design? The auto-routing feature in Altium ($8000) and all other EDA packages are shitty. Why can't Deep Learning help here?
Chase Walker
What would your training set look like? What would be its features? Do we need classification, clusterization or something else entirely? What you are describing is very vague.
Nolan Campbell
Sure, just press the magical "deep learn" button and the computer will do all your work.
This is totally how AI works and AI is not some stupid meme the media pushes because they have nothing better to report on.
Brody Murphy
>A ton of money is spent on Natural Language Processing and Speech Recognition and Synthesis to make shitty botnets and surveillance.
The main financier of natural language processing is the European Union.
They spend hundreds of millions on (human) translators every year, so they desperately want to automate that.
Owen Barnes
I am pretty sure some car companies are already working on something like this to improve the aerodynamics of their performance cars.
Austin Hughes
Let's take PCB routing as an example. It could automatically suggest trace-width, ideal location of vias, etc. There would already be a simulation that tests the circuit to so the ML algorithm could use this for classifying multiple existing layouts. It could also look at thousands of images of other PCBs of all sizes and complexities.
Here is a feature called ActiveRoute. It's a proprietary algorithm to generate multiple routes but it makes design mistakes often. Can't it be made context-aware by making a ML algorithm looks at existing PCB images and classify them by the number of layers? You can also take an X-Ray of PCBs to generate the training dataset.
Christopher Morgan
Forgot link - youtube.com
Zachary Reyes
Google, Amazon and Facebook spend billions.
Brayden Foster
No they don't.
Ryan Cruz
>et's say when you're designing something like a 4-axis robot, the CAD software could automatically include an adaptive stability design mod that a machine learning algorithm could come up with multiple options that you can quickly select and then continue working on it.
As someone who designed adaptive controllers for 6-axis arms, I can tell you this is complete gibberish. I suppose you could use ML for trajectory tracking instead of a traditional adaptive controller for some reason. I don't see how that provides any operational benefits over an adaptive control law. I'm sure some companies are doing this already, but more industrial applications are too scared to move away from PD controllers. I don't see what use it would be in the design of the arm either. You will end up spending more time training your network and formatting your inputs for a one time use if you're doing it in design.
As for auto routing PCBs, just use simulated annealing.
>muh machine learning will fix everything
is just a buzzword thrown around by people that are unaware of more efficient task oriented approaches that exist. It has its place but it isn't the end all be all
Jacob Wright
You are hugely over-estimating what AI can do.
Sure you can use it to optimize a design and automate some tedious tasks.
But don't expect any real intelligence, that's pure science fiction.
Adam Ward
I have a better idea. They should make an AI that can develop a better AI. Then the AIs will learn how to become an AI Engineer. When AI Engineers build a distributed AI network, it will be able to make better AIs.
Sebastian Flores
Good fucking luck. Deep learning is not an automatic solution to all your problems.
It can solve some specific problems very well, but to get it to solve new problems takes a great deal of work.
Christian Hughes
In order to use nearly any machine learning technique you need to reduce your problem down to a classification or a regression problem. I don't see a clear way of doing that here.
The only thing that may work is reinforcement learning, but your output space is of such monstrous dimentionality that I don't see it ever working. Reinforcement learning kinda sucks, and takes a lot of work and computational power to get it to do anything useful.
Jonathan Brooks
>he fell for the andrew ng meme
Christian Morgan
This is an example of a car chasis designed by an AI using an evolution algorithm (similar to the one on those silly games)
Christopher Nguyen
that looks retarded and difficult to manufacture