Learning with Chemical versus Electrical Synapses -- Does it Make a Difference?

TL;DR

The paper investigates whether chemical versus electrical synapses affect learning in bio-inspired neural networks when the same architectural framework is used. By comparing LTC-based chemical synapses and classic CT-RNN electrical synapses across sparse NCP and fully-connected wirings in a photorealistic lane-keeping task, it assesses performance, robustness, and interpretability using open- and closed-loop evaluations, along with saliency and neural-activity analyses. The results show that the wiring architecture substantially impacts performance and that chemical synapses provide notable gains in robustness and interpretability, with NCPs enhancing these benefits across both synaptic models. Practically, these findings guide the design of safe, interpretable, and reliable bio-inspired controllers for robotics and AI systems, indicating that combining sparse architectures with chemical synapses yields the strongest advantages.

Abstract

Bio-inspired neural networks have the potential to advance our understanding of neural computation and improve the state-of-the-art of AI systems. Bio-electrical synapses directly transmit neural signals, by enabling fast current flow between neurons. In contrast, bio-chemical synapses transmit neural signals indirectly, through neurotransmitters. Prior work showed that interpretable dynamics for complex robotic control, can be achieved by using chemical synapses, within a sparse, bio-inspired architecture, called Neural Circuit Policies (NCPs). However, a comparison of these two synaptic models, within the same architecture, remains an unexplored area. In this work we aim to determine the impact of using chemical synapses compared to electrical synapses, in both sparse and all-to-all connected networks. We conduct experiments with autonomous lane-keeping through a photorealistic autonomous driving simulator to evaluate their performance under diverse conditions and in the presence of noise. The experiments highlight the substantial influence of the architectural and synaptic-model choices, respectively. Our results show that employing chemical synapses yields noticeable improvements compared to electrical synapses, and that NCPs lead to better results in both synaptic models.

Figures (7)

• Figure 1: The 4-layer NCP stacked with a convolutional head. The ConvNet extracts information from the input image which then serves as sensory input for the NCP. In a fully connected architecture, we replace the NCP with all-to-all wirings.
• Figure 2: Structural-similarity index (SSIM) between saliency maps, under normal and Gaussian-noise conditions.
• Figure 3: Examples from the dataset: summer and winter
• Figure 4: Training and testing procedure.
• Figure 5: Summer-winter input images in (a), and their saliency maps of LTC NCP-19 in (b), LTC fully-19 in (c), CT-RNN NCP-19 in (d), CT-RNN NCP-64 in (e) and CT-RNN fully-64 in (f). Observe that an NCP architecture makes the models more focused compared to their fully connected version. The LTC NCP-19 keeps its attention on the road and its contours in summer, and on the road contours in winter. It ignores the irrelevant information from the border of the image.