Visualization of the NN
A screenshot of the demo tester in the app
When trying to think of a way to make a difference with the UMBC Psychology departments, we discovered this (relatively) simple idea that could really help parents protect their children.
What it does
The app listens for new text messages and sends them to our backend hosted on Google cloud. The server then sends the text into a neural net for processing. All insulting messages are then flagged.
How we built it
*Backend written in Go, deployed with docker. *The classifier was written in Python using Keras and Tensorflow using publicly available datasets. *App written in Ionic Framework.
Challenges we ran into
*Datasets are hard! A significant portion of the hackathon was spent looking for a working dataset and shaping the dataset to fit our needs. *Neural networks can do the complete opposite of what you want when the data is bad. *Learning how to use Ionic in a short amount of time. *Networking docker containers and debugging multiple services in the cloud is pretty daunting!
Accomplishments that we're proud of
*Working neural net with acceptable error rates for our purposes. *Robust distributed backend running in the cloud. *Good proof of concept app.
What we learned
Each of us wanted to tackle different parts of the project, and we all learned a lot about creating systems. Eric learned a lot about networking docker containers and web server dev, Brad learned a lot about Android dev through Ionic, and Nikita learned a lot about neural nets.
What's next for Cyber Bully Identifier
We have a lot of plans to expand the identifier. We found much better (much less callous) datasets online that we would like to train a more suitable network architecture on. As it turns out, training neural networks requires a lot of performance and LOTS of data. Many of the architectures we initially planned on testing were infeasible in the amount of time we had.
We also plan on working on the app much more. Developing the app was tedious but worth it. We want to implement the full functionality (along with some aesthetic tweaks).