NYRCS_numbers

This project will be used for an educational AI demonstration on NY research day (21st May 2025).

Trained on MNIST dataset obtained from https://www.kaggle.com/datasets/oddrationale/mnist-in-csv. https://drive.google.com/drive/folders/111m0ZpYKW9ADVDxhy1pkcCZowJjMmffv?usp=sharing. Download and unzip both csvs into this file.

Implementation

This project will let students draw a number and have an MLP decide what number it looks like.

It will be a flask webapp running locally on the computer at the station. The program is split into:

• process.py, an interface to call the AI. Use it as so:

from process import recognise
# process implements recognise(image: list) -> list
# Arguments:
# - image is a 28 x 28 2d array, where image[y][x] is a number from 0 to 255 indicating pixel darkness from darkest to lightest
# - image[0][0] is the top left corner of the image. Basically the conventional image format
# Return value: A list indicating [percentage probability it is 0, is 1, is 2, is 3, ...]

• ai.py implements all the ai math. The code is ugly because its all math 🙏
• evolution.py implements the evolution algorithm to generate a model. Refer to the # Training section for instructions on how to use it.
• Other file(s), to implement the flask frontend. For other team members to decide and update readme.

Training

Run evolution.py to generate a model through an evolution algorithm. You only need ~1000 generations. The script will automatically pick up where the last run ended.

Delete b1.npy, b2.npy, W1.npy, W2.npy and generations.txt files to restart. Copy over someone else's b1.npy, b2.npy, W1.npy, W2.npy and generations.txt files into your folder to use their model

TODO for this program:

1. Split work & implement flask frontend
2. (Francis) implement process.py