DiaLSM: Towards Write-Stall-Free Performance via Shard-based LSM-tree
为解决LSM树在持续写入压力下的写停滞问题,提出了一种基于分片的DiaLSM架构,通过将写-刷新-压缩路径拆分为多个独立分片来提高吞吐量并降低延迟。
为解决LSM树在持续写入压力下的写停滞问题,提出了一种基于分片的DiaLSM架构,通过将写-刷新-压缩路径拆分为多个独立分片来提高吞吐量并降低延迟。
研究提出ALPHABET模型,通过压缩时间序列到稳定复数极点模式并利用双银行架构处理特征轨迹,以少量参数实现高效且具竞争力的序列预测。
研究通过控制递归动态来解决测试时深度问题,提出了一种基于有限时间动态机制的方法以确保增加迭代次数不会降低答案质量。
This study investigates whether radar imagery can serve as an effective input modality for deep vision models to estimate air traffic complexity. The traffic situation is encoded into a five-channel image incorporating positional, heading, speed, and altitude information. For the first time, a Vision Transformer is employed to perform regression on four intrinsic complexity components derived from geometric relationships among aircraft. Experimental results demonstrate that the proposed approach achieves R² values exceeding 0.96 across all complexity components. Perturbation analysis further reveals the model’s sensitivity to key aircraft, with its responses closely aligning with those aircraft’s actual contributions to overall complexity, thereby validating the efficacy of sparse, self-similar radar images for modeling air traffic complexity.
This work addresses the challenge of learning parity functions—a canonical problem hindered by linear non-separability and global dependencies—by introducing TESLA, a novel activation function that explicitly controls polynomial order within the activation layer. Inspired by Fourier series, TESLA employs a learnable combination of sine and cosine components to selectively amplify high-frequency signals, thereby guiding the network to capture global structural patterns. Theoretical guarantees are provided through Lipschitz continuity and Rademacher complexity analyses. Empirically, TESLA achieves strong generalization on 32-dimensional parity tasks using only 100,000 samples—merely 0.002% of the input space—and demonstrates robustness to 30% label noise. It also significantly outperforms baseline methods on Forrelation and ImageNet-100 benchmarks.
为解决LSM树在持续写入压力下的写停滞问题,提出了一种基于分片的DiaLSM架构,通过将写-刷新-压缩路径拆分为多个独立分片来提高吞吐量并降低延迟。
研究提出ALPHABET模型,通过压缩时间序列到稳定复数极点模式并利用双银行架构处理特征轨迹,以少量参数实现高效且具竞争力的序列预测。
研究通过控制递归动态来解决测试时深度问题,提出了一种基于有限时间动态机制的方法以确保增加迭代次数不会降低答案质量。
This study investigates whether radar imagery can serve as an effective input modality for deep vision models to estimate air traffic complexity. The traffic situation is encoded into a five-channel image incorporating positional, heading, speed, and altitude information. For the first time, a Vision Transformer is employed to perform regression on four intrinsic complexity components derived from geometric relationships among aircraft. Experimental results demonstrate that the proposed approach achieves R² values exceeding 0.96 across all complexity components. Perturbation analysis further reveals the model’s sensitivity to key aircraft, with its responses closely aligning with those aircraft’s actual contributions to overall complexity, thereby validating the efficacy of sparse, self-similar radar images for modeling air traffic complexity.
This work addresses the challenge of learning parity functions—a canonical problem hindered by linear non-separability and global dependencies—by introducing TESLA, a novel activation function that explicitly controls polynomial order within the activation layer. Inspired by Fourier series, TESLA employs a learnable combination of sine and cosine components to selectively amplify high-frequency signals, thereby guiding the network to capture global structural patterns. Theoretical guarantees are provided through Lipschitz continuity and Rademacher complexity analyses. Empirically, TESLA achieves strong generalization on 32-dimensional parity tasks using only 100,000 samples—merely 0.002% of the input space—and demonstrates robustness to 30% label noise. It also significantly outperforms baseline methods on Forrelation and ImageNet-100 benchmarks.