Music 256a Reading Response, Fall 2021
Ray Gifford

Reading Response: Humans-in-the-Loop AI

After reading through and seeing examples of Humans-in-the-loop machine learning design, I like the idea. Although, I find the name itself confusing. "Humans-in-the-loop", to me, seems pretty intuitive on its own; but within the context of the many other terms in the world of machine learning... its difficult for me to place it. It's not clear to me if this approach can be used for both supervised and unsupervised ML algorithms... The internet says: "Bothhhh???". And also... how different is this from a reinforcement learning approach?

That aside, the idea of designing ML algorithms that respond and prompt for human interaction, seems nice. Further the idea of making something that goes beyond the strengths and capabilities of either piece on its own (human and computer), seems pretty ideal. In fact, it almost seems essential; if you are interested in making functional ML.. or at least functional with human acceptable precision. I could see this approach being especially useful when accuracy REALLY matters. But it may not always matter. I'm feel that plenty of business owners would be cool with automation of tasks that are average, or perhaps slightly below average that of a human completing the same task... but do not require paid employees to complete. However, any instance where accuracy is very consequential, like in manufacturing safety features or human-acceptable management, to me, that's when people should be concerned with the most accurate automation... and, perhaps they should consider HITL AI. Personally, I find interacting with Spotify or Apple Music frustrating, when they start asking me all of these questions about what I enjoy, so that it can better suggest music. "Just figure it out!" Lol. On the opposite extreme, my partner, who teaches music... finds that apple suggests so much music similar to the listening material that he assigns to his students, without capturing the music he is really into. There is an option to mark songs or albums or artists as disliked, for future calibration... maybe this is an example of HITL AI. But my partner does like listening to classical albums, it just doesn't have a good representation of what he enjoys in his free time most. So marking an album as disliked, doesn't help in his case. So, to me what seems more helpful, is a "supervised HITL AI" idk. Starting with a simple question at the beginning "Do you want me to check in with you to see if you like what I choose, or do you want me to just figure it?". That would be ideal, for me. Because I would really enjoy never seeing this occurring in the background, and would enjoy the automation regardless of accuracy. I'm sometimes pleasantly surprised by its suggestions, and that's fun for me.