UC Davis

The past decade has witnessed the rising dominance of deep learning (DL) and artificial intelligence (AI) in a wide range of applications. In particular, the ocean of wireless smart phones and IoT devices continue to fuel the tremendous growth of edge/cloud-based machine learning (ML) systems including image/video recognition and classification. To overcome the infrastructural barrier of limited network bandwidth in cloud ML, existing solutions have mainly relied on traditional compression codecs such as JPEG that were historically engineered for human-end users instead of ML algorithms. Traditional codecs do not necessarily preserve features important to ML algorithms under limited bandwidth, leading to potentially inferior performance. This dissertation investigates application-driven optimization of programmable commercial codec settings for networked learning tasks such as image classification.

In the first part of this dissertation, we focus on the efficient use and optimization of existing off-the-shelf commercial image compression codecs in bandwidth constrained image classification applications. We consider a cloud-based inference application where a power and memory limited embedded source device transmits the collected images to a powerful cloud server over bandlimited wireless channels. Our main contributions are two folds. Firstly, we show that the reconstruction step of the existing image decoders is unnecessary for cloud-based inference. Deep learning classifiers designed to take intermediate features as inputs, instead of RGB images, can perform inference few times faster with the same or improved classification accuracy. Secondly, we show that redesigning the entropy coders of commercial image codec such as JPEG2000 and learning optimal parameter setting of the entropy coders for a given task in end-to-end manner can significantly improve rate-accuracy performance of the codec.

In the second part, we investigate the methods of improving rate-distortion-accuracy performance in cloud-based AI applications for DL-based image compression codecs. Exploring end-to-end optimization of the complete codec, we propose novel classifier architectures based on variational auto-encoders (VAE) that outperform rate-classification accuracy of several conventional codecs. Further investigating DL-based codecs, we discuss how to achieve better rate-distortion-accuracy performance with end-to-end training revisiting the concept of region of interest (ROI).

In the third part of this dissertation, we explore recent interpretable information theory based concepts when modeling real world data and their applicability in data constrained deep learning scenarios. In particular, we investigate the use of linear discriminative representations (LDR) of images when designing cloud-based deep learning systems with improved rate- accuracy performance. Further, considering challenging but practical data constrained tasks such as zero-shot and few-shot learning, we investigate the generalization of such linear feature representations learned with rate reduction concepts.