PapersOfficialSelected85°00:16Chinese content从人类偏好中学习：安全AI新方法该方法解决了AI对齐中的核心难题——如何让AI理解人类真实意图，对构建可控AI系统具有里程碑意义。#ai-safety#reinforcement-learning#human-preferences#alignmentOPrimary sourceOpenAI BlogSourceRead laterReadWorth followingUsefulFollow ai-safety
PapersOfficial00:16Chinese content鲁棒对抗输入：多视角欺骗神经网络对AI安全研究和自动驾驶系统开发者极具警示意义，表明多模态感知并非对抗防御的万能药。#adversarial-attacks#robustness#computer-vision#autonomous-drivingOPrimary sourceOpenAI BlogSourceRead laterReadWorth followingUsefulFollow adversarial-attacks
PapersOfficialSelected14:00Chinese content生成对抗网络、逆强化学习和基于能量的模型的联系GAN、IRL、EBM 原来是一家#GAN#IRL#EBM#OpenAI2 sourcesEventOPrimary sourceOpenAI BlogSourceRead laterReadWorth followingUsefulFollow GAN
PapersOfficialSelected14:00Chinese contentWeight Normalization：一种加速深度神经网络训练的简单重参数化方法用简单重参数化加速训练，效果堪比BN但更轻量#Weight normalization#重参数化#训练加速#神经网络OPrimary sourceOpenAI BlogSourceRead laterReadWorth followingUsefulFollow Weight normalization