This paper explains the power of using multivariate embeddings with k-sparse competition (SP/TM in HTM language) to perform forecasting in noisy time series data. Sugihara’s work in this area was one of the main inspirations behind our theory of the Feynman Machine. We’ve recently written to him about the connections between his work and these alternative ways of looking at HTM-like systems.
Information leverage in interconnected ecosystems: Overcoming the curse of dimensionality Science, August 2016 [PDF]
Hao Ye, George Sugihara
Harnessing complexity in ecology
Ecology concerns the behavior of complex, dynamic, interconnected systems of populations, communities, and ecosystems over time. Yet ecological time series can be relatively short, owing to practical limits on study duration. Ye and Sugihara introduce an analytical approach called multiview embedding, which harnesses the complexity of short, noisy time series that are common in ecology and other disciplines such as economics. Using examples from published data sets, they show how this approach enhances the tractability of complex data from multiple interacting components and offers a way forward in ecological forecasting.
In ecological analysis, complexity has been regarded as an obstacle to overcome. Here we present a straightforward approach for addressing complexity in dynamic interconnected systems. We show that complexity, in the form of multiple interacting components, can actually be an asset for studying natural systems from temporal data. The central idea is that multidimensional time series enable system dynamics to be reconstructed from multiple viewpoints, and these viewpoints can be combined into a single model. We show how our approach, multiview embedding (MVE), can improve forecasts for simulated ecosystems and a mesocosm experiment. By leveraging complexity, MVE is particularly effective for overcoming the limitations of short and noisy time series and should be highly relevant for many areas of science.