Comprehensive Attention Benchmark (CAB)
收藏资源简介:
Comprehensive Attention Benchmark (CAB) 是一个综合性的注意力机制评估基准,由上海人工智能实验室开发。该数据集包含七个真实世界的任务,涉及计算机视觉、自然语言处理、语音处理和时间序列预测等多个研究领域。CAB旨在通过这些任务评估不同注意力机制在长序列建模中的效率和效果,特别是针对四种不同的注意力模式:非因果自注意力、因果自注意力、非因果交叉注意力和因果交叉注意力。数据集通过GitHub公开,以便研究者可以访问并使用这些资源进行进一步的研究和开发。
Comprehensive Attention Benchmark (CAB) is a comprehensive evaluation benchmark for attention mechanisms, developed by the Shanghai AI Laboratory. This dataset encompasses seven real-world tasks across multiple research domains including computer vision, natural language processing, speech processing, and time series forecasting. CAB is designed to assess the efficiency and effectiveness of various attention mechanisms in long sequence modeling via these tasks, with a specific focus on four distinct attention patterns: non-causal self-attention, causal self-attention, non-causal cross-attention, and causal cross-attention. The dataset is publicly available on GitHub, enabling researchers to access and utilize these resources for further research and development.

- 1CAB: Comprehensive Attention Benchmarking on Long Sequence Modeling上海人工智能实验室 · 2023年



