Connectome for Drones (huggingface.co)

🤖 AI Summary
A new project has been unveiled that leverages the FlyWire v783 connectome to enhance drone navigation through a spiking neural network (SNN). This network comprises 138,639 neurons and over 15 million synapses, with a topology that specifies the connectivity and excitatory/inhibitory nature of these connections. Utilizing an efficient training method known as e-prop, the SNN is tasked with navigating a drone in the gym-pybullet-drones environment, initially focusing on altitude control before refining the command output for multiple rotor speeds. This development is significant for the AI and machine learning community as it showcases a novel approach to drone control by mimicking biological neural networks. The project’s architecture, learning methodologies, and control experiments—such as using shuffled and random connection variants—offer insights into how different configurations affect behavior. With successful tests indicating the training algorithms' efficiency—achieving high cosine similarity between e-prop outputs and traditional backpropagation through time (BPTT)—this initiative not only contributes to robotics but also enhances our understanding of neural network dynamics and learning in biologically-inspired systems.
Loading comments...
loading comments...