Preprint · In progress
AIENs: Adaptive Initializer Enhancer Networks
Replaces the Gaussian prior in a VAE's KL-divergence loss with adaptive kernel density estimation, cutting training time by 19% and improving synthetic-data fidelity by 9%.
The standard VAE assumes a Gaussian prior. AIENs uses adaptive kernel density estimation (AKDE) instead, with matching changes to the VAE architecture. Because the prior already captures the data’s pattern, the model learns faster (19% less training time, so less compute) and generates data with 9% higher fidelity.