Neuromorphic Time-Series Processing for Robotic Applications

Context

Robotic applications require the ability to perceive and interact with an ever-changing environment across multiple sensory modalities, such as vision, sound, and touch. At the same time, robots must operate autonomously without constant access to cloud computing resources. This is a challenging task given the limited computational and energy resources available within autonomous robotic systems. One promising approach to enable efficient robotic applications is neuromorphic computing, which draws inspiration from biological neural networks that are sparse, asynchronous, and energy-efficient. However, training and deploying such systems necessitates different methods than conventional deep learning approaches. The goal of this work is to explore, study and implement the architectures of the Spiking Neural Networks suited for the time-series processing in robotic applications using neuromorphic hardware.

Targets

• Research state-of-the-art Spiking Neural Network architectures for time-series processing tasks in Robotics

• Reproduce existing SNN architectures and adapt to a broader set of the time-series processing tasks

• Evaluate the SNNs on a range of time-series benchmarks on Neuromorphic accelerators like Intel Loihi 2

Requirements

• Experience with Machine Learning and Deep Learning frameworks like Pytorch, Tensorflow or JAX

• High python proficiency