AI/ML Research Intern at BMW on Neuro-Inspired Computing for Tactile Sensing

Why Oscillatory Neural Networks?

AI methods are one of the key drivers of today’s technological progress in robotics and human-vehicle interaction. For example, machine learning can be used for tactile sensory processing to robustify robotic manipulation by inferring object properties such as stiffness and object pose. However, conventional hardware determined by the Von Neumann Architecture (CPUs, GPUs, etc.) is lacking for the use of energy-efficient Edge AI Inference needed in environments such as intelligent autonomous vehicles. One limitation of the traditional Von Neumann architecture is the split of memory and computation into separate physical components. This has two main drawbacks: a) the overall computational power is bottlenecked by the data transfer speed to transfer memory. And b), energy is wasted on transferring the data physically between all the different components, which results in a lack of energy efficiency. Biological systems do not lack these constraints. For example, the human brain consumes only around 20 W of power with its 80 billion neurons, while achieving much more intelligent behavior than current autonomous systems.

Taking human cognition as an inspiration, neuromorphic computing provides a possible research pathway to tackle the Von Neumann architecture’s limitations and to enable the wide deployment of Edge AI Computing required for robotics and intelligent autonomous systems. One interesting aspect of human brain computing is the rhythmic firing of neurons, which can be described by mathematical oscillations. These oscillations can be observed with electroencephalograms (EEGs) and vary in amplitude and frequency. Effects such as synchronization in the collective dynamics of neurons might be key to understanding higher cognitive functions such as memory. With Oscillatory Neural Networks (ONNs), we are investigating a novel way of approaching computing. ONNs have a lot of potential because a) ONNs can be natively implemented on highly energy-efficient neuromorphic chips with simple electrical units that naturally generate periodic behavior. And b), ONNs have the potential to process time-series data with minimal or no preprocessing.

The second point is crucial for the research goals of the Phastrac project. With conventional methods, time-series data need to be discretized and processed in chunks. But with ONNs, there’s the potential to process the raw time signal with no discretization and/or preprocessing (with an analog circuit). This means that there is no analog-to-digital conversion, i.e. the whole computing pipeline from sensors to computing only consists of analog components.

Phastrac Project

This research is conducted as part of the EU research project Phastrac. The goal of Phastrac is to develop the next-generation computing architecture for AI applications with ONN-based chips. In total, the project has 4 partners: IBM, BMW, TUe, and PPKE. Phastrac is split up into 4 work packages and tackles the whole pipeline of the computing architecture development. For example, IBM leads the research on the low-level chip design with the materials and fabrication processes of an ONN-based chip with work package 1. BMW leads work package 4 and is focused on showcasing AI application demonstrators where we use oscillatory neural networks for the sensory signal processing. We collaborate closely with the university PPKE from Hungary, because they provide us with the ONN simulators and learning architectures as part of work package 3.