Papers00:33Chinese content对比分布匹配:离散扩散中的摊销序贯蒙特卡洛方法离散扩散模型在文本、图结构等离散数据生成中至关重要,但采样效率一直是瓶颈。这篇论文提出的对比分布匹配方法直接解决了这个问题,做生成模型研究的开发者值得关注。#discrete diffusion#contrastive learning#sequential Monte Carlo#sampling optimizationAK@_akhaliqSourceRead laterReadWorth followingUsefulFollow discrete diffusion
PapersSelected09:46Chinese contentPVMC:10倍加速深度状态空间模型训练的新方法做时间序列建模或状态空间模型研究的团队,PVMC让训练速度提升10倍且不牺牲精度,值得直接尝试。#training acceleration#sequential Monte Carlo#generative modelsaarXiv cs.AI@John-Joseph Brady et al. (3)SourceRead laterReadWorth followingUsefulFollow training acceleration
PapersSelected14:41Chinese contentURGE:无需梯度的扩散模型推理时引导新方法URGE解决了扩散模型推理时引导的计算瓶颈,做生成模型优化或采样加速的研究者可以直接用这个梯度无关的轻量方案,值得关注。#diffusion models#sequential Monte CarloaarXiv cs.LG@Lifu Wei et al. (4)SourceRead laterReadWorth followingUsefulFollow diffusion models