A Hierarchical Multi-scale Cortical Learning Algorithm for Time Series Forecasting

Cortical learning algorithm (CLA) is a time series prediction method that is designed based on the human neocortex. Although the CLA attempts to extract temporal dependencies of time series, it ignores the characteristics of the data and can't deal with the intricate temporal patterns within the sequence. Multi-scale information is crucial for modeling time series, but is not fully studied in the CLA. To this end, we propose a Hierarchical Multi-Scale Cortical Learning Algorithm (HMS-CLA), which extracts representations of different scales and adaptively learns temporal variations of time series at each time step. First, a hierarchical CLA architecture is designed to capture multi-scale information of input sequence, where each layer corresponds to each scale. Then, at each time step, the hierarchical CLA learns temporal context under various time scales and the predicted results from different temporal contexts are adaptively decoded to generate the final output. Extensive experiments results on benchmark datasets show that compared with the vanilla CLA and HMS-CLA can reduce prediction error by 14% for time series.