Grasshopper

algorithmic modeling for Rhino

Information

Crow

This is the support group for the Grasshopper plug-in
Crow - Artificial Neural Networks in Grasshopper.
Please post your questions / requests / remarks here

Get Crow here:
http://www.food4rhino.com/project/crow?etx
or here:
http://www.felbrich.com/projects/Crow/Crow.html

Members: 26
Latest Activity: on Wednesday

Discussion Forum

This group does not have any discussions yet.

Comment Wall

Comment

You need to be a member of Crow to add comments!

Comment by Benjamin Felbrich on October 19, 2016 at 7:34am

Hello Winsion,

so Crow is based on artificial neural networks. These networks consist of logical nodes that propagate information, are structured in layers and are (de-)activated upon (not) exceeding a certain threshold. Lots of machine learning or so called AI techniques make use of this paradigm. Even Google's AlphaGo uses these techniques, though in a tremendously more advanced level than Crow can provide.
There are numerous books out there about this topic, that has been investigated for over 70 years! However, I found this one very enlightening:
http://page.mi.fu-berlin.de/rojas/neural/neuron.pdf

It's  basic introduction into the logic behind these networks and some of the most important advances in their development.
If you want to know more about the modern utilization of ANN and have some coding experience, you could look into Tensorflow, Google's Deep Learning API. Google, however, mainly makes use of these AI things for image classification and feature detection in images and videos. Not very useful for design, yet, but who knows what the future brings ;)

Comment by Winsion Liang on October 15, 2016 at 12:37am

Hello , I feel interest on your Plugin , but how can I study the base to understand it ? Thanks

Comment by Benjamin Felbrich on October 5, 2016 at 3:23am

Hey Kacper,

thanks a lot! Super cool that you use Crow. I actually didn't even test the Backpropagation component with the MNIST data set myself. Feel free to keep posting your results, I bet it's very helpful for others as well.

About the computation speed: I will try to parallelize the backpropagation component in the next release as I did in the SOG component (probably with a Boost option), to hopefully speed up computation a bit (5 to 40% depending on network topology). However, since the engine is written in C#, we probably won't get crazy high speed. But let's see...

Comment by Kacper Radziszewski on October 4, 2016 at 5:37pm

Hi Benjamin!

Thanks for a Crow.

I have played with it today using MNIST data set.

Had some good results so far, still using small learning data set, because of the time it takes to compute.

Here are my results on intigers recognition:

Backpropagation Network settings:
3000 training cycles

3x sigmoid layer

10 neurons per layer each
learning rate of 1.0

Training set of 500 examples

During classifying 400 examples accuracy of 45%


With a bigger training set which is available I should have better result, but as a first try with NN I believe it works.

I'm posting my definition if anyone would like to take a look and give me any feedback.

mnist data set neural network test

Comment by Benjamin Felbrich on September 14, 2016 at 11:11am

Hello there fellow group members. This is to announce that Crow 0.2.1 was just released on Food4Rhino:

- Renamed n-dimensional SOMs to SOG (Self-Organizing Grid) to account for the calculation of higher dimensional Kohonen-Maps (conventional SOMs are limited to two dimensions)
- Added a Boost Options to SOG (right click on SOG engine to activate) to perform multi-core computation. WARNING: This only results in speed up, if neuron count >~1000 and SOG dimension is >2. Otherwise it might actually slow down computation
- Added a timing option to SOG to monitor time per cycle and total calculation time
- few minor bug fixes

Have fun playing with it. Don't hesitate to post questions / remarks here.

Best
Ben

 

Members (26)

 
 
 

Translate

Search Grasshopper

Photos

  • Add Photos
  • View All

© 2017   Created by Scott Davidson.   Powered by

Badges  |  Report an Issue  |  Terms of Service