Long Short-Term Memory (LSTM) neural networks, invented in 1997 (by Hochreiter and Schmidhuber) have enjoyed a renaissance due to the advent of Deep Learning. Since 2013, LSTM has been successfully employed for tasks such as speech recognition, machine translation and image captioning. This project will explore the use of LSTM in detecting motion patterns in time-series sensory data.
First, experience will be gained with a simple stick-figure simulator of moving animals (human or otherwise), chosen by the students and instructor. A wide selection of simulated scenarios will then be designed, with each capable of producing long time-series of basic proprioceptive data, such as joint angles, and rotational velocities.
Next, a Deep LSTM network will be implemented in Tensorflow. It will receive proprioceptive time series and classify them into motion categories, such as walking, galloping and trotting (for a horse), or walking, running, skipping and limping (for a human).
IMPORTANT: If you sign up for this project, please send a) your CV (including a transcript with all of your college grades, and b) a brief explanation of WHY you want to do this particular project to Prof. Keith Downing (email@example.com)