Neuromorphic computing aims at implementing the computational principles of biological neural networks on dedicated hardware devices. Its main goal, however, is not necessarily to replicate neural circuits in the most realistic manner, but rather to exploit some of their advantages. In this context, spiking neural networks (SNNs) are a critical methodological component in neuromorphic systems. SNNs model neural information processing on the level of single impulses, or spikes, which allows extremely sparse, and thus highly energy-efficient, computing in an asynchronous, event-like manner. Thus, SNNs only compute when spikes occur. In contrast, regular neural networks, which most of the recent AI models—including ChatGPT or diffusion models—are based on, compute in an always fully active way making them tremendously inefficient. Neuromorphic implementations, in turn, can reduce the energy consumption of e.g. big vision models down to the range of milliwatts and even lower. Moreover, due to their compactness and efficiency, neuromorphic solutions can be integrated into sensors directly, paving the way for truly smart cognitive sensor systems. It is these features that make neuromorphic computing a candidate for an extremely disruptive technology, bearing the potential to accelerate intelligent automotive systems far beyond the yet imaginable. Here, we highlight recent advances in neuromorphic computing and discuss their potential for present and future automotive applications.
Session: Poster |