Brain-Inspired Networks Learn to Identify and Avoid Obstacles, Offering New Robotics and Disease Insights

Neuroscience-inspired neural networks are at the forefront of bridging the gap between biological intelligence and technological innovation. These sophisticated models offer a powerful lens through which to understand complex brain functions, while simultaneously providing adaptive and efficient control mechanisms for robotics. This research introduces a novel synaptic learning rule, drawing inspiration from the precise synchronization of synaptic inputs to individual excitatory neurons within a feedforward spiking neural network. The network itself is designed with a layered structure, incorporating three excitatory layers and two feedback inhibitory layers, starting with low connection probabilities and weak synaptic weights, allowing for dynamic evolution.

Under an unsupervised learning paradigm, the network was exposed to various stimulus patterns, facilitating the dynamic evolution of both synaptic weights and connectivity over numerous training trials. A core focus of the investigation was to understand how these dynamics were influenced by the intensity of feedback inhibition. The study successfully identified critical conditions that allowed the network to achieve stable activity, a prerequisite for reliable function. A key finding emerged in the evaluation of the model's pattern separation efficacy—its ability to distinguish between similar input patterns—and its direct relationship to the network's dynamics. As lead author Faghihi notes in the paper, "The results highlight the critical role of feedback inhibition in both stabilizing the network and enhancing pattern separation."

Specifically, the research revealed that a balanced synchronization between excitatory and inhibitory populations is paramount for maximizing pattern separation efficacy. Beyond providing a fresh computational framework for deciphering how information is processed within neural systems, this model also sheds light on cognitive disorders characterized by impaired inhibition and pattern separation, such as autism and schizophrenia. To demonstrate its practical utility, the trained network was embedded within a simulated agent navigating a two-dimensional environment. In this scenario, the network was tasked with a practical application: identifying a previously trained stimulus as an obstacle and subsequently avoiding it. This work offers a promising framework for advancing cognitive robotics, paving the way for novel approaches that mimic natural intelligence and support the learning of intricate environmental patterns.