Computation in Biological Neural Networks: From Biological Inspiration to Computational Models

This project focuses on the computational principles of biological neural networks, aiming to bridge the gap between biological neuroscience and machine learning, and develop next-generation computational models inspired by biological systems. Core research directions include dendritic computation, spiking neural networks, neuromodulation regulation, and network topology, while exploring their application potential in fields such as neuromorphic computing, brain disease understanding, and next-generation AI architectures.

Although modern deep learning-based artificial neural networks (ANNs) have achieved success, they have fundamental differences from biological neural networks. Biological neurons have complex geometric shapes (dendrites, axons, etc.), diverse dynamic characteristics (spike timing, plasticity), and multi-scale organizational principles, which are ignored in standard ANNs. This omission raises the question: Are we missing key computational principles of biological systems, leading to efficiency bottlenecks in artificial systems?

The project focuses on the following key questions:

1. Dendritic Computation: How does the spatial structure of dendrites affect computing power? Can we design artificial neurons with dendritic structures?
2. Spiking Neural Networks (SNNs): Explore STDP learning rules, event-driven efficiency, and temporal information processing.
3. Neuromodulation and Global State: What is the counterpart of the global regulation mechanism of biological neuromodulators in artificial networks?
4. Network Topology and Community Structure: How do connection patterns of biological networks (small-world, community structure, etc.) affect their functions?

The project adopts multiple technical approaches:

• Computational Modeling: Hodgkin-Huxley model (detailed biophysics), integrate-and-fire model (simplified spikes), compartmental model (spatial structure).
• Machine Learning Integration: Implement differentiable bio-inspired modules using PyTorch/JAX, combine biological learning rules (Hebbian, STDP) with gradient descent, and evaluate performance on benchmark tasks.
• Data Analysis: Utilize public neuroscience datasets (electrophysiology, structural, and functional imaging data).

The project's outcomes are expected to be applied in:

1. Neuromorphic Computing: Adapt to neuromorphic hardware, improve energy efficiency, and be suitable for edge devices.
2. Brain Disease Understanding: Build realistic neural models to assist in the study of mechanisms of diseases such as epilepsy and Parkinson's.
3. Next-Generation AI Architecture: Provide alternative solutions to Transformer limitations (context length, inference cost), especially in continuous learning and few-shot adaptation.
4. Brain-Computer Interface (BCI): More accurately simulate neural computation to improve BCI for motor recovery or disease treatment.

The project faces the following challenges:

1. Complexity vs. Interpretability Trade-off: The more realistic the biological model, the harder it is to understand. How to retain key principles while ensuring analyzability?
2. Training Efficiency: Bio-inspired learning rules (e.g., STDP) are slower than backpropagation. How to accelerate them or combine them with gradient descent?
3. Validation Difficulty: How to verify that models capture key principles of biological systems? Close collaboration with experimental neuroscience is required.
4. Hardware Support: Features like asynchronous spikes and analog computation are inefficient on traditional hardware, requiring specialized hardware support.

The project is suitable for an open science model and requires interdisciplinary collaboration. Potential contributors include: neuroscience researchers (providing biological constraints and validation), machine learning engineers (implementing and optimizing models), theoretical physicists (analyzing network dynamics), and hardware engineers (exploring neuromorphic implementations). Open-source code and data sharing are crucial for accelerating progress.

This project represents an important direction for improving artificial computing systems inspired by biological neural networks, serving as a complement and extension to existing deep learning methods. The evolutionary wisdom of the biological brain contains ununderstood computational principles; studying these principles not only enables the development of better AI systems but also deepens our understanding of the essence of thinking. The value of the project lies in establishing a cross-disciplinary exploration framework to promote collaboration among researchers from different backgrounds. In today's rapid development of AI, such basic exploration is particularly important.