BioTensors-AI: A Spiking Neural Network Brain Based on STDP Learning

BioTensors-AI is a research preview project that implements a spiking neural network (SNN) brain architecture with STDP (Spike-Timing-Dependent Plasticity) learning capabilities. The project uses a hierarchical cortical structure with 1100 neurons, aiming to explore low-power, memory-first autonomous decision-making solutions for edge AI. It is not intended to replace deep learning but to provide a biologically more plausible alternative path for specific scenarios (such as real-time response and online learning).

Although traditional artificial neural networks (ANNs) have succeeded in tasks like image recognition, they have issues such as high energy consumption and reliance on large amounts of labeled data, limiting their application on edge devices. The human brain completes complex cognition with about 20 watts of power, thanks to features like event-driven processing, local learning, and sparse activation. Spiking neural networks (SNNs) are inspired by this, and BioTensors-AI further extends biological plausibility to a complete brain architecture, implementing a hierarchical processing flow of sensation-association-execution.

BioTensors-AI includes four core components:

1. BioTensor: A LIF-style spiking neuron layer that supports adaptive thresholds, fatigue mechanisms, homeostatic regulation, and STDP learning (if the presynaptic neuron fires first, synapses are strengthened (LTP); otherwise, they are weakened (LTD)), no backpropagation required.
2. BioBrain: A hierarchical cortical architecture including a sensory layer (400 neurons), association layer (400), execution layer (100), limbic layer (100), and Broca layer (100), enabling recursive processing and functional division of labor.
3. BioBrainPro: Multi-modal extension supporting three types of input: health sensing, navigation sensing, and voice sensing.
4. BioCompiler: A training/compilation pipeline responsible for concept imprinting, ethical shielding (implanting safety constraints), and weight export (.bio format).

The current portable prototype uses approximately 1100 neurons with an input dimension of 100 spike dimensions. The project is implemented based on Python 3.10+ and NumPy, with minimal dependencies for easy expansion. Additionally, a Tkinter-based visual monitoring tool is provided to observe network activity in real time.

BioTensors-AI has significant advantages for edge autonomous decision-making scenarios with low CPU and memory-first requirements:

• Aerospace Applications: Real-time response to unexpected situations (e.g., obstacle avoidance) in deep space exploration, with low power consumption adapting to spacecraft energy budgets.
• Autonomous Robots: Efficient processing of sensor inputs in dynamic environments, with STDP learning supporting continuous environmental adaptation.
• Health Monitoring: Long-term operation on wearable devices, real-time analysis of physiological signals, and abnormal alarm triggering.

Key differences between BioTensors-AI (SNN) and traditional deep learning:

Feature	Traditional Deep Learning	BioTensors-AI (SNN)
Computing Mode	Dense Matrix Operation	Sparse Event-Driven
Learning Algorithm	Backpropagation	STDP (Local Plasticity)
Energy Consumption	High (GPU-intensive)	Low (Sparse Spiking)
Data Requirement	Large Amounts of Labeled Data	Online Learning/Small Number of Samples
Biological Plausibility	Low	High
Hardware Friendliness	Requires Specialized Accelerators	Natively Supported by Neuromorphic Chips

BioTensors-AI is currently in the early research stage (v0.1.0). The portable prototype contains 1100 neurons, and larger-scale training is conducted on an independent workstation. It may be gradually open-sourced in the future. The project author Christophe is a Belgian pharmacist and independent developer with an interdisciplinary background integrating neuroscience, pharmacology, and computer science, bringing a unique perspective.

BioTensors-AI is an interesting exploration in the field of neuromorphic computing, providing a biologically plausible alternative for scenarios requiring low power consumption, real-time response, and online learning. With the development of neuromorphic hardware such as Intel Loihi and IBM TrueNorth, its practicality will improve. The project uses the MIT license, and contributors interested in neuromorphic computing, onboard autonomy, robotics, or aerospace applications are welcome to participate.