Is visual rendering the future abstraction for optimizing quality of experience (QoE)? A new workshop paper (2008.04128 A New Abstraction for Internet QoE Optimization) out of the University of Chicago and Microsoft Research seems to think so.
Why change the abstraction? Although user-perceived QoE is subjective by nature, Internet video streaming optimization has typically centered on optimizing a QoE model based on human engineered features. And optimizing QoE has largely plateaued, despite an ever growing list of features.
Visual rendering, on the other hand, is “a video stream that records all of the pixels displayed on the screen over time as seen by the user.” This provides two main advantages:
- expressibility: it captures a full visual experience of a user, thereby bypassing the need for human engineered features and potential abstraction mismatch with actual user perception
- generalizability: it is a unifying abstraction for all Internet applications
What sort of architecture would be required for this new approach? The paper outlines two main requirements:
- A visual renderer that infers the visual rendering of taking an optimization action.
- A QoE model that takes the visual rendering as input and predicts the QoE of a given optimization action.
It then goes on to draw upon ideas from other disciplines to suggest ideas for how QoE modeling based on visual rendering might work. From cognitive visual perception, it borrows the ideas of expectation and attention to posit high QoE as meeting exceptions without the attention region. From computer vision, visual attention detection, video summarization, and video prediction techniques could be used in implementing the new QoE model concept.
Will this be a new frontier in machine learning in networking? I guess we’ll have to wait and see.