Training diffusion neural networks by backpropagating differentiable rewards
EP4762488A1Pending Publication Date: 2026-06-24GDM HOLDING LLC
View PDF 0 Cites 0 Cited by
Patent Information
- Application Number
- EP2024786711
- Authority / Receiving Office
- EP · EP
- Patent Type
- Applications
- Current Assignee / Owner
- Priority Date
- 2023-09-28
- Filing Date
- 2024-09-30
- Publication Date
- 2026-06-24
AI Technical Summary
Technical Problem
Diffusion neural networks struggle to align their training data distribution with desired post-training behavior, leading to a mismatch between modeled and desired output qualities.
Method used
The system employs a differentiable reward function to fine-tune the diffusion neural network through reinforcement learning, allowing it to generate data items with desired properties by backpropagating rewards through the sampling process.
Benefits of technology
This approach enables the diffusion neural network to produce outputs that meet specific quality criteria, such as aesthetic appeal, while reducing computational and memory costs through techniques like gradient checkpointing and truncated gradient computation.
✦ Generated by Eureka AI based on patent content.
Smart Images
Figure EP2024077489_03042025_PF_FP_ABST
Abstract
Methods, systems, and apparatus, including computer programs encoded on computer storage media, for training a diffusion neural network using a differentiable reward function.
Need to check novelty before this filing date? Find Prior Art