Advancing Spiking Neural Networks: The NeuEvo Framework

By Dr. Sampath LonkaReviewed by Susha Cheriyedath, M.Sc.Sep 26 2023

In a recent publication in the journal Proceedings of the National Academy of Sciences, researchers introduced an innovative approach to enhance spiking neural networks (SNNs) by incorporating diverse neural circuits inspired by biological nervous systems. These circuits encompass feedforward and feedback connections and excitatory and inhibitory neuron types, resulting in a richer SNN architecture.

Study: Advancing Spiking Neural Networks: The NeuEvo Framework. Image credit: LariBat/Shutterstock

Background

Human cognition, encompassing perception, reasoning, and decision-making, arises from a dynamic combination of neurons, circuits, and brain regions. Conventional artificial neural networks fall short of mimicking the human brain's structure and computation, often relying on single neural circuits and real-valued units.

Spiking neural networks (SNNs) emulate the brain's information processing using asynchronous spikes. Unlike real-valued artificial neural networks (ANNs), spiking neurons exhibit rich spatiotemporal dynamics, processing spatial input, accumulating temporal information in membrane potential, and firing spikes upon threshold crossing.

Capitalizing on SNN characteristics, they excel in adversarial resilience, noise robustness, continuous learning, and simulating cognitive functions. However, compared to deep ANNs, they lag in performance for tasks such as image classification and object detection. Current high-performance SNN structures merely replace nonlinear activation functions in nonspiking ANNs while inheriting their complex architectures, risking information loss due to sparse transmission.

The NeuEvo framework

Researchers introduce the NeuEvo strategy, drawing inspiration from the brain's evolutionary process. NeuEvo employs unsupervised spike-timing-dependent plasticity (STDP) learning for network structure refinement and a global error signal for weight adjustments. The NeuEvo framework involves the spiking neurons model, data encoding, and construction of the neural circuit evolution and optimization.

For the spike-neuron model, researchers used the Leaky Integrate-and-Fire (LIF) neuron model in discrete time. Static data encoding in SNNs, especially for images, involves rate coding and direct coding. Rate coding converts real-valued images into spike trains, while direct coding transforms input currents into spike sequences, significantly reducing simulation time.

NeuEvo incorporates excitatory and inhibitory neural layers with different receptive field sizes and dense connections between initial neuronal clusters, fostering diverse neural circuit formation. Neurons receive various neuronal inputs based on synaptic weights, influencing membrane potentials and spike transmissions and enhancing network diversity.

Neural Circuit Evolution involves assigning weights to the influence of local neural activity and global error signals, evolving neural circuits over time. Initially, multiple circuits are established, and as the network learns from the environment, it adaptively selects appropriate circuits for information transmission, allowing unused circuits to degrade. STDP updates circuit selection weights, striking a balance between various circuits. Additional mechanisms prevent excessive excitation and facilitate competition among neurons.

Simultaneous optimization of global and local weights is considered bilevel optimization. To efficiently optimize, an alternating method updates synaptic weights and neural circuit selection weights iteratively. This approach facilitates the simultaneous optimization of synaptic weights and neural circuit structures, mimicking the evolution of neural structures in response to external stimuli.

Results and analysis

The authors first presented visualizations of the evolved neural circuits and their corresponding spiking activities. The network's performance is assessed in both classification and reinforcement learning tasks.

Visualization of Evolved Neural Circuits: NeuEvo's capacity to autonomously construct neural circuits is showcased on the DVS-CIFAR10 dataset, a dynamic vision sensor (DVS) version of the computer vision dataset CIFAR10. These circuits embody diverse connection types among neurons, mirroring the complexity of the biological nervous system. These automatically generated neural circuits can be deconstructed into distinct basic neural circuits, each serving specific functions.

Furthermore, an analysis of 20 different neural circuit structures generated with varied random seeds on the DVS-CIFAR10 dataset reveals the prevalence of feedforward excitatory circuits. The balance between feedforward and lateral inhibition is nearly equal, while feedback and mutual inhibition circuits are less frequent.

Performance in Reinforcement Learning: NeuEvo's versatility and efficiency are demonstrated as the evolved neural circuits are employed in reinforcement learning tasks. These circuits exhibit strong generalization capabilities and can be effectively integrated with both off-policy and on-policy deep reinforcement learning algorithms, yielding competitive results compared to artificial neural networks (ANNs). The six simulated robotic tasks in OpenAI Gym underscore the ability of NeuEvo-derived neural circuits to transfer effectively to tasks with high-dimensional observation and control spaces, closely resembling real-world scenarios.

Performance in Classification: NeuEvo's effectiveness in image recognition tasks is evaluated across various datasets, including static and event datasets. The results are compared with other network structures, both manually designed and automatically searched. In static datasets, the inclusion of inhibitory neurons significantly enhances performance, while feedback connections prove more beneficial in event datasets due to their temporal correlation.

Additionally, NeuEvo's STDP module's effectiveness for evolving neural circuits is compared with randomly sampled neural circuits. The results demonstrate NeuEvo's superiority in constructing efficient SNN architectures, regardless of the search space.

Energy Efficiency Analysis: SNNs, characterized by spike-based information transfer, inherently offer higher energy efficiency compared to nonspiking ANNs. The relationship between accuracy and the number of spikes in SNNs is explored, with NeuEvo-generated models consistently requiring fewer spikes to achieve superior performance. This reflects the high energy efficiency of NeuEvo-derived SNNs, a vital aspect of neuromorphic hardware with a network-on-chip (NoC) architecture.

Conclusion

In summary, researchers introduced the NeuEvo framework, drawing inspiration from the brain's evolutionary process. Validation of the framework was carried out through classification and reinforcement learning tasks. Experimental results demonstrate the superior performance of SNNs constructed using NeuEvo. Future research will explore the influence of different neuron types and synaptic plasticity rules on neural circuits. The goal is to further refine and advance the development of biologically plausible neural networks based on insights from neural system evolution.