Circuitry with Neural Networks

Inspiration

We wanted to visualize a neural network with circuitry. Initially, we wanted to build a physical neural network (PNN) that conducts computations using conductance and resistance, with inspiration from https://arxiv.org/abs/2406.03372. However, we pivoted to modeling a neural network using LEDs similar to The Stillwell Brain by Vsauce.

What it does

The CwNN classifies a drawn digit on the computer and lights up the corresponding activated neurons in the LED neural network. Layers are color-coded, and users can hover over activated neurons in specific layers to see which neurons in the previous layers caused it to activate. This enables users to visualize the process of a neural network better.

How we built it

The CwNN was built using LED strips that were soldered together, an Arduino UNO, and JavaScript for the frontend.

Challenges we ran into

It was difficult working with hardware because our team had little experience with circuitry. Additionally, indexing the LEDs was tedious since there were hundreds of LEDs.

Accomplishments that we're proud of

We're proud that we were able to connect the frontend on a computer to send signals to the Arduino to dynamically change the lights.

What we learned

We learned how to send data from the computer to the Arduino and how to model a neural network with only discrete weights.

What's next for Circuitry with Neural Networks

We would like to implement CwNNs with back propagation and to train our own neural network. Additionally, instead of visualizing a neural network, we would like to implement a physical neural network through current.