We love the speed of swipe typing on mobile keyboards, but too often the predictive keyboards put words in our mouth.
What it does
Swiper-Go-Swiping re-envisions the mobile keyboard as a circle to improve accuracy on the already speedy swipe text feature.
How we built it
By having the keyboard in a circle, the swipe patterns for each word is more distinctive. We've trained a model to recognize the unique swipe pattern for each word. We used Google Cloud Platform to train our model.
Challenges we ran into
We had to create our own data set. We drew the set by hand, converted it into digital files, created xml files for each image, converted the xmls to csv, and converted the csv files to tfrecords.
Accomplishments that we're proud of
This is our first try at machine learning. For a while, it's been our team's goal to do a hack with machine learning, and we made it happen!
What we learned
We learned about gc storage buckets, and tfrecords, and labeling data, and training data, and testing data.
What's next for Swiper-Go-Swiping
We would love to design and code the actual mobile interface. Ideally, we would put the keyboard on our own devices to create training data by using the keyboard and recording the patterns for each word.