Overcoming Real-world Deployment Challenges for AI-enabled Systems

Abstract

With recent breakthroughs in artificial intelligence (AI) research and innovation, AI promises the potential to dramatically improve human quality of life. However, these technologies often fail to reach widespread adoption due to unsolved real-world deployment challenges. Some examples include AI algorithms that drain the host device's battery too quickly and systems that exhibit significant performance degradation when utilizing a low-cost, noisy sensor. In this dissertation, I motivate the need to devote more effort to identifying and investigating implementation challenges. I then argue that the creative application of ML techniques like transfer learning, adversarial learning, knowledge distillation, and autoencoding can make these systems more efficient, effective, and scalable, leading to wider adoption. I demonstrate evidence for this thesis through four projects: (1) efficient cough detection for mobile phones, (2) cougher identification with transfer learning, (3) induction of emotional prosody in speech synthesis, (4) BP estimation with blood volume signals on smartphones.