NLP Practice Project

Inspiration

We wanted to learn more about machine learning, so we wanted to create a very small scale ML application to dip our toes into the field.

What it does

The application looks at a 28x28 pixel image of a white alphabet letter on black background. It sends it into a ML model trained on EMNIST letter data, which then attempts to guess the letter in the image.

How we built it

We imported TensorFlow to serve as the library from which all of our machine learning code would come from. From there, we constructed a rudimentary AI trained on pictures of alphabet letters from the EMNIST library. Then, we fed the AI a 28x28 picture of an alphabet letter to see if it would determine the correct letter.

Challenges

Since we were completely new at machine learning, we utilized a lot of external resources to get the project to the point where it is. We used tutorial videos on Machine Learning from 3Blue1Brown, and referenced the TensorFlow documentation as well. David Ramirez of General Dynamics Missions Systems played a big role in advising us on our project- without his advice, it is likely that we wouldn't be anywhere near where we are with our project now.

Accomplishments:

We as a group are very proud that the application actually ran and sometimes correctly identified the provided input as the correct letter. We also gained enough bare intuition about the model, so that we could change the various parameters to improve the model's accuracy at letter detection. We feel that we had genuine moments of insight into machine learning and its applications.

What we learned:

This was our first collaborative project. We learned how to work with a team, and the struggles of dividing work without a collaborative platform like replit. In our research before the hack, we learned about NLP- Natural Language Processing. Since this was our first foray into ML, we also obviously learned ML and AI, as well as gained a healthy respect for what goes into making and training an AI.

What's next for NLP Practice Project:

Now that we have a machine learning model that can identify letters, our primary goal is to improve accuracy to over 99%. Then, with a solid base of letter identification, we can work to have it iteratively be able to recognize larger groups of letters: words, sentences, paragraphs. Also, we would want to make it so that it can correctly identify letters in non-ideal conditions regarding things like light, spacing, and alignment.