The closest I have seen an "AI" (or modestly sized neural net) "predict" was in a guitar-to-MIDI translator box. At the lowest fundamental frequencies, waiting for a whole (let alone several, for improved accuracy) waveform periods to be captured takes too long to be perceived as instantaneous (the MIDI communication and synthesizer played will add further delay), and the signal of a freshly attacked string doesn't look that clean either. But NN was able to reasonably reliably detect the vibration frequency with a fraction of a period captured - which some might want to call predicting the future (continuation of not very sinusoidal waveform whose shapes vary wildly with how the string is attacked and what it did before). One might say that's just pattern recognition, but I guess there is overlap between that and "predicting the future".
Great content, very useful, thanks!
The closest I have seen an "AI" (or modestly sized neural net) "predict" was in a guitar-to-MIDI translator box. At the lowest fundamental frequencies, waiting for a whole (let alone several, for improved accuracy) waveform periods to be captured takes too long to be perceived as instantaneous (the MIDI communication and synthesizer played will add further delay), and the signal of a freshly attacked string doesn't look that clean either. But NN was able to reasonably reliably detect the vibration frequency with a fraction of a period captured - which some might want to call predicting the future (continuation of not very sinusoidal waveform whose shapes vary wildly with how the string is attacked and what it did before). One might say that's just pattern recognition, but I guess there is overlap between that and "predicting the future".
What types of models were being used? Is it everything from ARIMA to NNs?