CIDARLAB / 3DuF

Interactive microfluidic design editor
http://3duf.org
BSD 2-Clause "Simplified" License
75 stars 71 forks source link

Fluid as alternative to Mathematical model of Artificial neuron #192

Closed imvetri closed 4 years ago

imvetri commented 4 years ago

Is your feature request related to a problem? Please describe. I'm always frustrated when modern Artificial intelligence uses neurons to train instead of fluids

Describe the solution you'd like Pressure to a fluid is the impulse to a neuron, shape and the path of a fluid's container is a memory.

Describe alternatives you've considered Use fluid design as an alternative to mathematical model of artificial neurons

Additional context I tried to contact CIDAR after watching this video. https://www.youtube.com/watch?v=aHvfEOlh_b4. I couldn't find any email to send this message to, instead opening an issue to communicate what great potential your lab's work is in the modern world of Artificial intelligence.

As a hobby I had been writing my thoughts out in form of a readme doc in github related to artificial intelligence. As days went by, I found convincing that fluids can replace artificial neurons in machine leaning and my main motivation is to get rid of math involved in modern coding.

My most recent writing is this https://github.com/imvetri/artificial-intelligence/blob/master/Self.Observing.model.md. Since it doesn't communicate clearly the thoughts I had, I made a video reading through the page with more details to it. Here is the audio recorded with video https://vimeo.com/user101685652.

I'm looking forward to hear from you, Thanks, Vetrivel.

rkrishnasanka commented 4 years ago

Hi @imvetri I'm closing this issue because its not related to 3DuF. However feel free to message me on Twitter @rkrishnasanka , I can point you to a couple of papers where microfluidics were used to solve a computational problem. My initial thoughts are : Approach 1 - You need to formulate whatever optimization problem you setup with machine learning as a fluidic problem. The challenge here would be to translate the entire optimization algorithm into fluidic primitives. Approach 2 - You want to model the neurons (non-linearity) using the fluid dynamics, so you can look at the shape / form of the final final output as the solution. I believe Photonic Neural Networks might have similar topologies (I can be wrong here). For this approach the challenge would be to convert the digital inputs into a fluidic signal and the fluidic output back into a digital format.

imvetri commented 4 years ago

Hi @rkrishnasanka. Thanks, the explanations greatly helped.