Memory in Memory: A Predictive Neural Network for Learning Higher-Order Non-Stationarity From Spatiotemporal Dynamics

By Yunbo Wang, Jianjin Zhang, Hongyu Zhu, Mingsheng Long, Jianmin Wang, Philip S. Yu. Computer Vision and Pattern Recognition, 2018

422 citations recorded by the source corpus.

Natural spatiotemporal processes can be highly non-stationary in many ways, e.g. the low-level non-stationarity such as spatial correlations or temporal dependencies of local pixel values; and the high-level variations such as the accumulation, deformation or dissipation of radar echoes in precipitation forecasting. From Cramer's Decomposition, any non-stationary process can be decomposed into deterministic, time-variant polynomials, plus a zero-mean stochastic term. By applying differencing operations appropriately, we may turn time-variant polynomials into a constant, making the deterministic component predictable. However, most previous recurrent neural networks for spatiotemporal prediction do not use the differential signals effectively, and their relatively simple state transition functions prevent them from learning too complicated variations in spacetime. We propose the Memory In Memory (MIM) networks and corresponding recurrent blocks for this purpose. The MIM blocks exploit the differential signals between adjacent recurrent states to model the non-stationary and approximately stationary properties in spatiotemporal dynamics with two cascaded, self-renewed memory modules. By stacking multiple MIM blocks, we could potentially handle higher-order non-stationarity. The MIM networks achieve the state-of-the-art results on four spatiotemporal prediction tasks across both synthetic and real-world datasets. We believe that the general idea of this work can be potentially applied to other time-series forecasting tasks.

Subfields: Atmospheric Science, Environmental Engineering, Global and Planetary Change

View Paper: https://www.semanticscholar.org/paper/16f9bfac4a68cc99c38c06db2aa523bc6537db0d

How researchers see this paper

The researchers whose trusted networks rank it highest, strongest advocate first. Every researcher gets a different order.

In Atmospheric Science

1. Philip S. Yu
2. Zhifeng Gao
3. Nicolas Thomé
4. Kevin Trebing
5. Jianmin Wang
6. Yunbo Wang
7. Siamak Mehrkanoon
8. Mingsheng Long

In Environmental Engineering

1. Philip S. Yu
2. Mingsheng Long

In Global and Planetary Change

1. Philip S. Yu
2. Zhifeng Gao
3. Kevin Trebing
4. Jianmin Wang
5. Yunbo Wang
6. Siamak Mehrkanoon
7. Mingsheng Long

Citations in the corpus

• Self-Attention ConvLSTM for Spatiotemporal Prediction (2020, methodology): “To enhance the ability of PredRNN on modeling high-order dynamics, Memory in Memory (MIM) (Wang et al. 2019) introduces more memory cells to process non-stationary and stationary information, which achieves the current SOTA performance in spatiotemporal prediction while result in a multiplication…”
• PredRNN: A Recurrent Neural Network for Spatiotemporal Predictive Learning (2021, methodology): “we use the MIM model [4], which is also based on the STLSTM unit, to specifically show the generality of the memory decoupling loss to different network backbones.”
• Disentangling Physical Dynamics From Unknown Factors for Unsupervised Video Prediction (2020, methodology): “Following the setting of [14, 11, 10], we predict 4 future frames given 4 input frames.”
• SmaAt-UNet: Precipitation Nowcasting using a Small Attention-UNet Architecture (2020, methodology): “The memory function and memory flow of this model were optimized in another implementation that added stacked memory modules [18].”