What about neural networks?

[Lorenzo Bolzani]

Hi, I recently started to work with neural networks and I was wondering if you are considering this technology for parts of this project.

I'm sure most of you already knows how it works but to put it in the context of this project it may be the possibility to simulate any complex non-linear function that may be much more complex to do in other ways.

You provide the input and the expected output and let the "learning algorithm" (typically some kind of gradient descent) to figure out the rest, reconfiguring its own "weights" to later be able to provide correct outputs for new inputs.

It may be a very accurate spike response model for example. While the training may take a lot of time the execution is extremely fast.

Let me know if it may be of interest to investigate something in this sense.

Bye

Lorenzo

[Tom Portegys]

Hi Lorenzo,

When I first came upon OpenWorm a few years ago, coming from a machine learning artificial neural network viewpoint, it also occurred to me to try to learn how the worm's sensory-motor network might be trained to do some input-output tasks. As interesting as that project was, of course the worm's neurons are not perceptron-like and don't learn as an ANN does. It remains a huge challenge to discover how the worm's connectome is weighted to allow it to function. There are a number of synaptic and non-synaptic neurotransmitters to deal with. For such a relatively simple creature, it is amazingly complex!

Tom

--
Visit us online at:
http://openworm.org
http://blog.openworm.org
http://github.com/openworm
http://twitter.com/openworm
https://plus.google.com/s/openworm
---
You received this message because you are subscribed to the Google Groups "OpenWorm-discuss" group.
To unsubscribe from this group and stop receiving emails from it, send an email to openworm-discuss+unsubscribe@googlegroups.com.
To post to this group, send email to openworm-discuss@googlegroups.com.
Visit this group at https://groups.google.com/group/openworm-discuss.
To view this discussion on the web visit https://groups.google.com/d/msgid/openworm-discuss/CAMgOLLz59k7Ux5-9whqk7%2BPjNzBkB0OaVC1-N_NS%2BJJHj8VWDA%40mail.gmail.com.
For more options, visit https://groups.google.com/d/optout.

[Lorenzo Bolzani]

Hi Tom,

what I was thinking about wasn't to model the whole neural system with an NN, but rather to use the "NN technology" to model/simulate a few arbitrary parts. You could probably accurately simulate the whole worm external behavior with a NN but this, as you say, wouldn't be of much help for the biological insights.

A single neuron response may be modeled with a deep NN comprising thousands of artificial "neurons", maybe with a "recurrent" architecture. Or a single sodium-potassium pump may be another example.

I gave a quick glance at spike-response models some time ago and I think that the general approach is to look for an analytical description. While this has many advantages I think it limits the possibility to really "fit"/replicate the actual response with all the real world subtleties that a compact mathematical representation has to drop.

You'd just replace a low level "black box" function (chemical, electrical, etc.) with another one. If the extra accuracy is relevant or not is something I've absolutely no idea about.

Lorenzo

To view this discussion on the web visit https://groups.google.com/d/msgid/openworm-discuss/CABcandPq_LDMbXQVzR8%3D%2BFDZOBHP03uD%2B_NCjxUMNpiVSq%2BswQ%40mail.gmail.com.

[Tom Portegys]

Hi Lorenzo,

Others in OpenWorm are better suited to commenting on the state of modeling individual neurons, but using a black box deep NN to produce high fidelity input-output values does sound interesting!

Tom

To view this discussion on the web visit https://groups.google.com/d/msgid/openworm-discuss/CAMgOLLyiJYsV%3Dj6xgMZHyvLbTu--iX_h1x%2BTK70_O6h1qc2Ejw%40mail.gmail.com.

[Stephen]

These links may be relevant to the questions here:

http://docs.openworm.org/en/latest/faq/#openworm-simulation-and-modeling

http://docs.openworm.org/en/latest/modeling/#neurons

https://github.com/openworm/CElegansNeuroML/blob/master/CElegans/pythonScripts/c302/README.md

https://github.com/openworm/muscle_model/blob/master/README.md

https://github.com/openworm/simple-C-elegans/blob/master/notebooks/simple-C-elegans.ipynb

Best,

  Stephen