Hyper-Connections
This addresses a fundamental architectural issue in deep learning for AI researchers and practitioners, though it appears incremental as an alternative to existing connection methods.
The paper tackles the problem of drawbacks in residual connection variants like gradient vanishing and representation collapse by introducing hyper-connections, which allow dynamic adjustment of connection strengths and layer rearrangement, resulting in significant performance improvements in large language model pre-training and vision tasks.
We present hyper-connections, a simple yet effective method that can serve as an alternative to residual connections. This approach specifically addresses common drawbacks observed in residual connection variants, such as the seesaw effect between gradient vanishing and representation collapse. Theoretically, hyper-connections allow the network to adjust the strength of connections between features at different depths and dynamically rearrange layers. We conduct experiments focusing on the pre-training of large language models, including dense and sparse models, where hyper-connections show significant performance improvements over residual connections. Additional experiments conducted on vision tasks also demonstrate similar improvements. We anticipate that this method will be broadly applicable and beneficial across a wide range of AI problems.