我们研究了频域MLPs学习到的模式,并发现了它们具有的两个固有特性,即(i)全局视角:频谱使MLPs能够全面理解信号并更轻松地学习全局依赖关系;(ii)能量压缩:频域MLPs专注于较小的频率成分,使信号能量更加集中。接着,我们提出了FreTS,这是一种简单而有效的架构,基于频域MLPs进行时间序列预测。FreTS主要包括两个阶段:(i)...
SoS Certificates for Sparse Singular Values and Their Applications: Robust Statistics, Subspace Distortion, and More 31 p. Two-component Dark Matter and low scale Thermal Leptogenesis 128 p. Vector-like quark doublets, weak-basis invariants and CP violation 5 p. Distributed Mixture-of-Agents...
Official implementation of the paper "Frequency-domain MLPs are More Effective Learners in Time Series Forecasting" - aikunyi/FreTS