NeuroPhone: An Open-Source Project for Running Neuro-Symbolic AI on Mobile Phones

NeuroPhone is an open-source Android application project aimed at fully deploying neuro-symbolic AI on mobile devices. It innovatively integrates Spiking Neural Networks (SNNs) with Large Language Models (LLMs), with all computations done locally (unless cloud fallback is actively enabled), balancing privacy protection and low-latency on-device intelligence. This article will cover background, architecture, technical highlights, and other aspects—welcome to discuss~

Cloud AI has issues such as privacy leakage risks (data uploaded to servers) and high latency due to network dependency (round-trip time over 500ms). NeuroPhone addresses these pain points through on-device local processing: sensor data does not need to be uploaded (unless cloud fallback is selected), and neural states are stored in the local directory ~/.local/share/neurophone/ to maximize user privacy protection.

NeuroPhone's architecture consists of six modules:

1. Sensor Layer: Reads data from accelerometers/gyroscopes, extracts features at 50Hz after IIR filtering;
2. Liquid State Machine (LSM): 512 neurons in an 8×8×8 grid, processes time-series signals at 1kHz frequency;
3. Echo State Network (ESN): Reservoir with 300 neurons (spectral radius 0.95), responsible for state prediction;
4. Bridge Layer: Integrates neural states and converts them into natural language context—core of neuro-symbolic conversion;
5. LLM Layer: Integrates Llama3.2 (1B/3B parameters, Q4_K_M quantization, 700MB size), performs local inference via llama.cpp;
6. Cloud Fallback: Optional Claude API call, supports hybrid inference decisions.

Technical highlights of NeuroPhone:

• On-Device Intelligence: All neural network processing is done locally, maximizing privacy;
• Low Latency: Inference latency <100ms, far better than cloud solutions;
• AI-Assisted Installation: Voice commands can automatically complete device detection, dependency installation, model download, and other processes;
• Modular Rust Architecture: Clear crate hierarchy (lsm/esn/bridge etc.), easy to extend and maintain.

Use Cases:

1. Privacy-first AI assistant;
2. Offline-capable intelligence;
3. Context-aware (sensor fusion to understand the environment);
4. Real-time response scenarios. Significance: Provides a new paradigm for mobile AI, demonstrates the feasibility of neuro-symbolic AI on devices, and has reference value for researchers in privacy/offline capabilities/new neural architectures.

Current Limitations: Targeted at technical users, requires Termux usage experience. Future Directions:

• Support more Android devices and chips;
• Expand sensors (heart rate/GPS);
• Enrich the pre-trained model ecosystem;
• Optimize the graphical interface to lower the entry barrier.