Excited to announce DPO has gone multi-modal! New paper out on RLHF for text-to-image diffusion models! We obtain large-scale state of the art results with 70% win rates against Stable Diffusion XL on human evals! Deep dive below 🧵
Our paper "Sparse Probabilistic Circuits via Pruning and Growing" was selected as Oral at
#NeurIPS2022
. Check here , or at poster session 3
#904
.
Can we scale up probabilistic circuits simply by increasing model sizes? 1/2
We propose a model pruning technique leveraging the probabilistic semantics of parameters; our approach can prune 97% of the model parameters without hurting performance, and based on which we are able to effectively learn large yet sparse probabilistic circuits. 3/3
We show that PC training performance plateaus as model size increases; most capacity in existing large PC structures is wasted: fully-connected parameter layers are only sparsely used. 2/3
@SCarbain
You are right. We train the model from the paired dataset end-to-end. One benefit of this method is that we don’t require a reward function e.g. PickScore.