PlaceField2BVec: A bionic geospatial location encoding method for hierarchical temporal memory model

• A new geospatial location encoding method, PlaceField2Bvec, is proposed for HTM model. • PlaceField2BVec enables HTM neurons to form a place field like a place cell. • The distribution patterns of the place field of HTM are consistent with those in the rat brain. • The PlaceField2BVec has the highest accuracy after enough training in extended geospatial space. Encoding geospatial location is a fundamental problem for geospatial artificial intelligence (GeoAI) research. In recent years, some methods (such as Place2Vec, Space2Vec, and Sphere2Vec) were proposed to encode geospatial point as a high-dimensional vector. However, all these geospatial location encoders were designed to generate a real number vector. So, when applied to some of the brain-inspired neural networks, such as Hierarchical Temporal Memory (HTM), which required the input of a binary vector, the existing methods failed. To solve the problem, based on the research from neuroscience about place cell, we proposed a new geospatial location encoding method called PlaceField2BVec. The method used the place field model to encode a location. The place field was represented by the summation of four Gaussian functions, allowing it to be stretched or divided into multiple fields as the geospatial space expanded. Then we created an HTM and devised an experiment that simulated rats moving on tables of varying sizes. The moving trajectories were encoded by PlaceField2BVec and input to the HTM. After training, we found that the artificial neurons of HTM formed a place field similar to those of hippocampal neurons in the rat brain and the distribution patterns of the place field from the two kinds of neurons were consistent. At last, our method was compared with existing Space2BVec and Buffer2BVec in terms of location prediction accuracy and to demonstrate the robustness of the binary vector encoding methods, two brain-inspired artificial neural networks— HTM and BinaryLSTM were used. The result showed that, for HTM, in smaller geospatial space the PlaceField2BVec and Buffer2BVec had about the same accuracy on average but the highest accuracy of PlaceField2BVec is 100 %; when the geospatial space extended, our method had the highest accuracy and the average accuracy of PlaceField2BVec, Space2BVec, and Buffer2BVec is 83.9 %, 25.2 % and 69.7 % after 20 times’ training. For BinaryLSTM, PlaceField2BVec always had the highest accuracy in location prediction although the accuracy decreased as the space extended. Our research can be utilized for machine self-localization, navigation, and location-related GeoAI applications, and it also contributes to the theory of cognitive maps.