Instead of just teaching Deaf people how to get along with the rest of the world and understand them, why don’t we make an effort to understand them and their language via technology?! In this talk we will learn how to use machine learning to interpret sign language.
Gestures: composed of hand posture, motion, and position with respect to each other, body parts, objects, etc.
A bit about signed languages:
India has a shortage of sign language interpreters - only about 250 certified people in the entire country, for millions of d/Deaf people.
There are parallels between signed and spoken languages in terms of machine recognition. Both are full natural languages and have all the complexities of such. However, signed language research didn’t start until the 1960’s, so it is newer.
Signed languages have some aspects spoken languages don’t have, such as facial expressions, signing speed as a modifier of verb meaning, and… well, three dimensions. You need to deal with lighting, differences in pose and camera angle, etc. Various attempts to solve this with multiple cameras, gloves, etc. have been tried. There have also been various attempts to develop annotation systems that transcribe signed languages and otherwise represent them in 2D writing.
Convolutional neural networks (CNNs) are a type of feed-forward neural network inspired by animal vision. As data flows from one layer to the next, it is progressively analyzed at higher and higher levels. Recurrent neural networks (RNN) resemble lists, and also cascade data forward. Exploding and vanishing gradients are an issue that RNNs amplify because they layer so many things atop each other.
This project used a RNN model with Long Short Term Memory (LSTM) units. CNNs are used for spatial features, and RNNs for temporal ones.
This was the end of the talk - the speaker offered to take questions outside or over Slack.