Pinglab: Exploring Excitation-Inhibition Dynamics and PING Mechanism in Spiking Neural Networks

Pinglab is an open-source project focused on Spiking Neural Network (SNN) research. It implements PING (Pyramidal-Interneuron Gamma) dynamics by explicitly modeling excitatory and inhibitory neuron populations, uses surrogate gradient methods for training, and proposes the "Δt Stability" metric to verify the model's generalization ability in continuous-time dynamics, providing a research platform for biologically inspired neural computing.

As the third generation of neural networks, Spiking Neural Networks (SNNs) have advantages in energy efficiency and temporal information processing, but there are core challenges in training: traditional backpropagation is difficult to directly apply to discrete spike signals. Surrogate gradient methods provide a solution to this problem. Pinglab was born in this context; it is not only an SNN implementation but also a platform for studying excitation-inhibition dynamics and PING mechanisms, aiming to explore biologically inspired network dynamics to improve generalization and stability.

Pinglab adopts the excitatory/inhibitory separation principle in neuroscience, based on the biological observation that brain neurons are divided into excitatory (glutamatergic) and inhibitory (GABAergic) types. In implementation, excitatory neurons transmit positive signals to enhance activity, while inhibitory neurons provide negative feedback to regulate excitability, allowing the network to spontaneously generate temporal dynamic patterns such as oscillations and synchronization, which are difficult to capture by traditional artificial neural networks.

The PING mechanism describes the generation of oscillations in the gamma frequency band (30-100Hz): excitatory pyramidal neurons activate inhibitory interneurons, which then inhibit pyramidal neurons to form a periodic cycle. Pinglab introduces this mechanism to enable the network to generate rhythmic activities similar to the brain, enhancing its expressive ability. This may be related to cognitive functions such as attention and memory consolidation, helping to explore the transformation of biologically inspired computing principles into improved machine learning performance.

Due to the non-differentiability of spike functions, direct backpropagation is difficult. Pinglab uses the surrogate gradient method: discrete spike activation functions are used in forward propagation, and smooth surrogate functions are used in backpropagation to calculate gradients. This strategy supports end-to-end training and allows the selection of multiple surrogate gradient functions, providing a flexible experimental platform for studying SNN training dynamics.

Pinglab proposes the "Δt Stability" metric, whose core idea is: a model that understands continuous-time dynamics should maintain stable performance under different integration time steps. The test process is: after training with a fixed step size, evaluate under different step sizes. This metric checks whether the model has learned the essence of continuous time rather than overfitting to specific discretization patterns, which is crucial for the time-precise computation of SNNs.

Pinglab helps understand the connection between biological and artificial neural networks and provides an experimental platform for biologically inspired computing principles. Application prospects include: edge computing/neuromorphic chips (event-driven to adapt to low power consumption), temporal signal processing (audio, EEG, etc.), and brain-like computing (building systems close to biological intelligence). Δt Stability provides a new dimension for SNN evaluation, making up for the deficiencies of traditional metrics.

Pinglab is open-sourced under the MIT license with a clear code structure and complete documentation. Participation suggestions: first read the style guide to understand the specifications; the project provides a unit test suite covering core functions, with slow test markers for selective execution. Researchers can explore the impact of biological mechanisms on network learning and generalization by experimenting with different E/I ratios, PING parameters, etc.