Paradigm Shifting Research on AI Output (in: Machine Learning and…)
By: Gamma UT
19 Aug 07:33
Bit of a flaw in the old methodology there if you want to research how changes in data during training affect it.
Model ablation is an old technique and has its uses but in something as opaque as an LLM or diffusion where the inputs can affect many, many weights, it has limited application. They also devised a different type of model that let them do the ablation in a targeted way that appears to generate the same kinds of results as regular diffusion models.
Go to first unread post in this topic
Read more here: Source link