A custom neural net for recoginzing handwritten digits, with a serious nod to I Am Trask. The neural net is pre-trained on the MNIST dataset (run on an AWS instance). And the user input is rescaled and tested.
sketch.js is a library created by Michael Bleigh to allow the user to draw the image of a digit. It has been modified to pull a drawn image from the page before being resized and sent to the neural network for classifaciton.
- Analize Results
- Analize correlation of misclassification and actual numbers (7's tend to be classified as 2, etc.)
- Train net on rotated images for more robust recognition
- Visualizations of each layer of the net (in d3 or bokeh)
- Provide a selection of neural net architectures to allow a user to swap out and see the different vizualations and results. (Results data will be stored by specific architecture.)
Navigate to: Finnegan
The baseline architecture for the net is on Github, here, and the code for the webapp (including Finnegan) be found here.
Full documentation hosted here
To adjust hyper-parameters or choose dataset: Open net_launch.py and adjust as needed at the bottom of the file. The dataset can be switched by commenting out the appropriate line at the bottom and uncommneting the other. Or simply:
from Network import network
And feed it the appropriate parameters.