AI isn’t “stable”, at least not in the way I’m used to thinking of computer operations. Or data.
To make progress, I need to accept that there isn’t a “final” version of the data I’m working on, given my budget of time and money. Not 100% happy about it, but accepting it as part of the attributes of this tool.
Circumplex
- Final and stable “Primary” emotions are the core criteria. For “Secondary” emotions, although the vast majority are stable, there are certain songs that show variation, and I’m accepting this variation with guardrails. My criteria for these is that they stay within the same “family” or theme, in this case:
- Alert, Excited, Happy
- Contented, Relaxed, Calm
- Bored, Depressed, Sad
- Angry, Distressed, Tense
- I’m also accepting that certain songs will vary in whether they show one or two secondary emotions. I could demand that there are always 2 Secondary emotions per song, but this isn’t true of every song and will lead to false results in a different direction. Or I could “micro manage” Chat for this data set, but as I’d like to reuse this Modular Expert on other song sets, I don’t want to fine tune it to much to this set. In essence trading reusability for perfection.