Skip to content

HiPPO

High-order Polynomial Projection Operators

Gu A, Dao T, Ermon S, et al. Hippo: Recurrent memory with optimal polynomial projections[J]. Advances in neural information processing systems, 2020, 33: 1474-1487.

在线函数逼近(Online Function Approximation)

RNN

c(t0)=[0.11.13.72.5]c(t1)=[1.52.90.32.0]

HiPPO 的目标是:在任意时间点 t,对输入信号 f(t) 的历史信息进行最优低维逼近。它通过正交多项式基函数(如 Legendre 多项式)的投影操作,动态更新系数以捕捉信号的变化趋势。

HiPPO-LegS (Legendre with Scaled Measure)

HiPPO-LegT (Legendre with Translated Measure)