Machine Learning for N-Dimensional Spatial Reasoning Tasks on the Web

Blake Moody^*JieHyun KimSanghyuk KimDaniel Haehn

• University of Massachusetts Boston, Boston, United States

Spatial reasoning is essential for solving complex tasks in dynamic and high-dimensional environments. However, current training models for spatial tasks are computationally demanding and heavily reliant on human input. To address this gap, we present Snake-ML, a web-based simulation tool and proof-of-concept framework designed to demonstrate client-side training of spatial reasoning tasks. Snake-ML serves as an efficient and intuitive test bed for developing spatial navigation strategies in browser-based environments. We chose the Snake game as our test bed because it is well-suited for demonstrating spatial reasoning in low-dimensional visual spaces while remaining relevant to higher-dimensional tasks, compared to alternative methods. Through quantitative analysis, on the edge alone, Snake-ML achieves a 4.58x speedup in model inference. Additionally, we developed a direct TensorFlow.js GPU pipeline that achieves up to 32x speedup in training time without any CPU/GPU synchronization. This pipeline has the potential to improve many edge-based AI visualization projects. Snake-ML shows potential for adaptability to complex spatial tasks, such as autonomous systems, robotics, and AI-driven environments. Our code1 and web-based simulation tool2 are publicly available.