arXiv:2406.16971v1 Announce Sort: new
Summary: Normalizing flows are a versatile class of chance distributions, expressed as transformations of a easy base distribution. A limitation of ordinary normalizing flows is representing distributions with heavy tails, which come up in purposes to each density estimation and variational inference. A preferred present answer to this downside is to make use of a heavy tailed base distribution. Examples embody the tail adaptive circulate (TAF) strategies of Laszkiewicz et al (2022). We argue this could result in poor efficiency because of the issue of optimising neural networks, reminiscent of normalizing flows, beneath heavy tailed enter. This downside is demonstrated in our paper. We suggest an alternate: use a Gaussian base distribution and a last transformation layer which may produce heavy tails. We name this method tail remodel circulate (TTF). Experimental outcomes present this method outperforms present strategies, particularly when the goal distribution has giant dimension or tail weight.
Supply hyperlink