We’ll be at DCC 2018 to present some of our current work on information storage in next-generation semiconductor devices – come say hi! Here is a summary of our submission which you can read in full on our publications page:
The traditional paradigm of image compression focuses on designing source coders that process images into a string of bits. These bits are then taken by a channel coder that either transmits or stores them on a binary channel. In this work, we are addressing a novel setting for image compression in which the data is to be stored on an analog device: Phase Change Memory (PCM). We are interested in these devices because of their benefits over traditional binary storage devices, chiefly: power-consumption, speed, and endurance. Instead of hand-designing analog channel coders for storing images on these devices — often an arduous task — we propose an adaptive autoencoder framework that accomplishes joint source-channel coding. We find that our proposed joint source-channel coding scheme is able to achieve a rate-distortion performance that is superior to that achieved by JPEG combined with a binary channel coder. In principle, our method can adapt to the statistics of a broad range of data types and memory devices, potentially even adapting to changes in device properties over time.