# Multi-Scale Representation Learning for Spatial Feature Distributions using Grid Cells

**Authors:** Gengchen Mai, Krzysztof Janowicz, Bo Yan, Rui Zhu, Ling Cai, Ni Lao  
**arXiv:** [2003.00824](https://arxiv.org/abs/2003.00824)  
**Date:** 2020-02-16  
**Summary page:** https://www.summarizepaper.com/en/paper/2003.00824v1/

## Key points

- Unsupervised text encoding models have advanced NLP by using neural networks to convert words into vector space representations based on positions and contexts within sentences.
- Spatial analysis aims to incorporate absolute positions and spatial contexts of geographic objects into models.
- Currently, there is no general-purpose representation model for space beyond simple discretization or feed-forward nets applied to coordinates.
- Inspired by grid cells in mammals, the Space2Vec model encodes absolute positions and spatial relationships of places.
- Space2Vec outperforms established machine learning approaches like RBF kernels, multi-layer feed-forward nets, and tile embedding approaches for location modeling and image classification tasks.
- Detailed analysis shows that Space2Vec's multi-scale representation excels at handling distributions at different scales.
- Space2Vec provides a valuable solution for encoding the absolute positions and spatial relationships of places.

## Summary

Unsupervised text encoding models have made significant advancements in Natural Language Processing (NLP) by using neural networks to convert words into vector space representations based on their positions and contexts within sentences. Similarly, spatial analysis aims to incorporate absolute positions and spatial contexts of geographic objects, such as Points of Interest (POIs), into models. However, there is currently no general-purpose representation model for space beyond simple discretization or feed-forward nets applied to coordinates. Inspired by Nobel Prize-winning Neuroscience research on grid cells in mammals, which provide a multi-scale periodic representation critical for location encoding and place recognition, we propose a representation learning model called Space2Vec. This model encodes the absolute positions and spatial relationships of places. To evaluate its effectiveness, we conduct experiments on two real-world geographic datasets for two different tasks: predicting types of POIs based on their positions and context, and image classification leveraging geo-locations. The results demonstrate that Space2Vec outperforms well-established machine learning approaches like RBF kernels, multi-layer feed-forward nets, and tile embedding approaches for both location modeling and image classification tasks. Detailed analysis reveals that while all baselines can handle distribution at one scale reasonably well, they exhibit poor performance in other scales. In contrast, Space2Vec's multi-scale representation excels at handling distributions at different scales. In conclusion, our proposed Space2Vec model provides a valuable solution for encoding the absolute positions and spatial relationships of places. Its multi-scale representations offer superior performance compared to existing ML approaches when it comes to location modeling and image classification tasks.