Early Work

In 2013, I presented my research at ISMIR, introducing possibly the first formal framework using multi-dimensional feature-space analysis to objectively benchmark the output distribution of generative systems against human-authored corpora. The core contribution was a novel statistical method for comparing AI-generated output to human-created work.

By mapping generative outputs into a shared geometric space, this work brought quantitative methods to the study of model bias and creative variance, moving beyond subjective Turing-style judgment toward statistical verification and evaluation. The same comparative logic (measuring synthetic output distributions against original human distributions) is what modern model evaluation systems formalize at scale.

This work was developed independently of Word2Vec (Mikolov et al., 2013), published the same year, which introduced learned feature embedding spaces to the mainstream. It predates the Inception Score (Salimans et al., 2016) and Fréchet Inception Distance (Heusel et al., 2017) for images, and Yang & Lerch's adaptation of embedding-based evaluation for music (2018) — the same comparative logic, arrived at independently, years earlier, with the tools of the time.