Inspiration
Modern ML models are growing larger in parameters, requiring more computing power to run. This takes away from some possible use cases in embedded devices, such as self-driving cars, as it's very difficult and costly to run the model locally. We looked into ways of running ML models on low-powered embedded systems with improved performance. As we wanted to challenge ourselves to run an ML model on a very low-powered device, we tried to implement a CNN model software onto a Nintendo 3DS.
What it does
We developed a 3DS homebrew software to run a CNN model to predict what number was drawn on the touchscreen in real-time, highlighting the massive performance improvement on a very low-powered device from more than a decade ago.
How we built it
We used the open-source DevKitPro toolset to be able to compile C code for the Nintendo 3DS. We also used the ncnn library to optimize the ML model for embedded devices. All our code was written in C/C++ and compiled for the 3DS.
Challenges we ran into
Getting everything to work cohesively (The 3DS has quite the set of quirks and features)
Accomplishments that we're proud of
Getting it to work (sort of)
Built With
- 3ds
- ai
- c++
- cnn
- homebrew
- ml
- nintendo
Log in or sign up for Devpost to join the conversation.