AI In Training, Part 2: Tapping the Well
In our last post, we discussed the importance of quality metadata when training Machine Learning and Artificial Intelligence. Today, we follow onto that with discussion on the practical uses of AI, and info on how we at TTA craft quality human metadata for ML Model training.
Four months out from our last post on AI, it seems that many of our predictions are holding true.
Namely, that despite relentless hype, massive capital expenditures, missed deadlines, and countless examples of LLMs underperforming in practice, AI products continue to have more to prove to the world of creators than not.
Also, with the increasing scarcity of training data and expanding glut of content deals with publishers to secure the rights to it, the general tone of the AI industry seems to have tempered greatly:
- Maybe AI shouldn’t replace your ability to write letters to your loved ones and heroes, but instead be useful relegated to toning down your work emails a bit
- If we can’t create an AGI super-intelligence that solves all human problems immediately, maybe we start by making one that’s really good at math
- Instead of GPT-5 inevitably coming to replace every human with a brain, maybe we instead focus on affordable models that do small things well enough to be helpful
Despite what might sound like criticism, TagTeam Analysis has long stood at the intersection of cutting-edge technology and human perception, be that with our laser-focus on SEO-specific music-search tags, or with the proprietary tagging tools we’ve been using in-house for over a decade to turn out quality, consistent music metadata at scale.
It’s our opinion that this latest shift in expectations for AI may prove to be hugely important for the maturation of AI’s role in many industries, including Production Music.
So let’s dive in!
SCALING THE DATA WALL
We’ve mentioned it before, but it’s not surprising that many modern AI products are starting to look more and more like search. That’s because in many cases, both AI and search do essentially the same thing: make sense of large swaths of information publicly available on the internet.
For practical purposes, both are truly revolutionary technologies. Especially in the more modest fields of Computer Vision and Machine Learning, well-trained AIs with datasets specialized for the purpose are already helping doctors detect cancer cells in tissue samples, at a scale that would otherwise be completely impractical for trained doctors.
In Production Music, we’ve seen similar attempts to separate signal from noise in large datasets, with companies like AIMS and Cyanite providing enterprise-level solutions for large catalogs needing basic searchability at a scale that would otherwise be difficult or costly to achieve.
But much like with SEO tags or search metadata, the quality and diversity of the training dataset for a Machine Learning model will directly impact its efficacy, accuracy, and flexibility across different sources.
Imbalances in training data can prove particularly troublesome when training accurate ML Models, something usually only solved by better sampling of accurate and representative data. (And if you need help solving this problem in your own data, don’t be shy.)
And while a Google search of an obscure topic with scant online sources may come up empty, ML inference on source material it hasn’t encountered before will often prompt a blind guess, resulting in detected material getting pigeonholed into another category entirely, and usually in ways that are hard to detect.
AI ACCOUNTABILITY
This last issue is, in our opinion, the crux of what makes many AI products weak in the marketplace today. Because in highly competitive industries (like Audio Post Production and Film/TV Advertising), high accuracy is often just as important as high accountability.
If a human assistant were hired to help find music for trailers and only did so accurately 80% of the time (which is about what we see from the most competitive ML models today), that assistant may be asked to study up over the weekend and show up to work the next week better suited for the task.
But if that human assistant then came back to perform that same lackluster work every day, you would imagine their employers might start second-guessing if it was worth having them around at all.
Hallucination, or the return of inaccurate or made-up information, continues to be an unsolved problem in AI, even in the most advanced models. ChatGPT goes so far as to warn users not to trust it:
With many researchers suggesting that AI hallucination will continue to be part of the way these models work, it’s clear that the only path towards utility for this technology will necessarily include human oversight and input.
The negative impacts of relegating responsibility to a machine that can’t detect its own mistakes has already proven disastrous, notwithstanding the issue of whether we want AI technologies replacing human jobs wholesale at all.
The alternative—AI technology coupled with the appropriate human oversight—we believe has a brighter future.
HUMAN EXPERTISE
This is where we can actually start to see AI as a good thing.
Because as AI grows in capability and specialization, and the limitations of the technology become better acknowledged, we believe human intelligence will actually increase in value as a necessary safeguard on a technology that will likely remain flawed for years to come.
In our view, human expertise (much like that of our analysts highly trained in music categorization) will remain indispensable, both for making sure AI tools continue to be accurate, and necessary for the ongoing training and updating of models to keep them relevant as culture changes over time. (Our work with ShutterStock and Epidemic Sound has given us a window into how this expertise will continue to be valuable to those looking to make the most of this powerful technology.)
Our business has worked for over a decade to make the highly technological process of digital music discovery human and real.
AI, for all its oversized promises and threats, appears to be the next frontier in that pursuit.
For more information on the metadata services we provide, feel free to contact us to discuss your project.