A convolutional neural network implemented on a Teensy 3.6 Microcontroller

A working implementation of a convolutional neural network written in Arduino C for Teensy 3.6.

The code in this repository was written for an interactive art piece. It is not a reusable template for your own project, sorry! I'm open sourcing it nonetheless hoping it might be helpful to someone.

Thanks to Yann LeCun for guidance and to Alfredo Canziani for explaining the math to me!

Architecture:

0   input     8x8 px
1   conv.     4x kernel size=3, stride=1, pad=0
2   maxpool   4x 2x2 px ReLu
3   linear
4   output

Usage

The neural net implementation itself is in neuralnet.ino: nn_predict() computes the gradient for a single sample. nn_train() completes a single training step.

It includes the necessary functions like nn_conv() for a forward pass, nn_conv_backwards() to calculate gradient descent loss, and nn_update_conv() to update weights.

Note: Initial weights are pre-generated in Network training.ipynb and hardcoded.

See diagram.ino for general operating and order of commands. The file includes other code: LED displays for layer outputs as well as motorized potentiometers that encode the network's weights.