Notable

Inspiration

Taking notes by hand increases material retention--this has been proven. But it also increases something else: the chance of losing your work. What if you could have the learning benefits of handwriting notes but still be able to keep a copy as a Google or Word document and Ctrl-F through it later? We had to tackle this problem ourselves.

What it does

Scribr is a deep learning model that allows you to input pictures of your notes and have it transcribed for you.

How we built it

We trained data from the IAM Handwriting Database on Tensorflow in the cloud. We used a CNN -> LTSM -> CTC structure.

Challenges we ran into

After bricking our computers trying to download all the data, we decided to move our data aggregation and model training to Google Cloud Platform’s Cloud ML Engine. This allowed us much more time for optimizing our model.

Accomplishments that we are proud of

Figuring out how to integrate Google Cloud Platform into our workflow was a lifesaver. Our app would not be where it is without it.

What we learned

We learned a ton about Convolutional Neural Networks and Long Short Term Memory Networks while building our project.

What's next for Scribr

There's still room to improve our model through more data and better architecture, which is going to be vital going forward. We also want to create a Flask app to serve our model.