Amazon Science·· 27 天前AI 评分56
AWS 提出 Priming 方法,无需从零预训练即可将 Transformer 改造为混合架构
Can we change an AI model’s architecture without retraining it from scratch?
AI 导读
Amazon Web Services 团队在论文《Priming: Hybrid state space models from pre-trained Transformers》中提出 priming 方法,复用预训练 Transformer 权重构造混合架构,把此类架构的预训练成本降低超过 99%。
来源:Amazon Science · amazon.science