To research, design, implement, and evaluate an ultra-low-power Spiking Neural Network (SNN) architecture that leverages in-memory computing principles for efficient online learning.Background:
The field of neuromorphic computing seems to offer a transformative solution for achieving intelligence at the edge. By emulating the brain's efficient biological mechanisms through Spiking Neural Networks (SNNs)
, neuromorphic computing systems not only promise substantial energy efficiency but also enhance real-time processing capabilities when integrated with online learning.
The conventional von Neumann computing architectures, characterized by separate memory and processing units, encounter performance constraints due to the continual data transfer between these segments. This structure leads to heightened energy consumption and processing time. Additionally, the widespread reliance on energy-intensive dynamic random-access memory (DRAM) exacerbates these energy concerns, particularly when grappling with the intensive computational requirements of online learning tasks in SNNs. In response to these challenges, the research landscape is shifting. Notable innovations like IBM's TrueNorth chip, which mirrors neural networks, are emerging. Alongside these digital solutions, there's a burgeoning interest in exploring analog, hybrid, and advanced nanoelectronic devices, with a keen focus on those boasting memristive attributes. In-memory computing, which conducts calculations directly within memory storage, has become a popular design choice, further reducing energy while decreasing latency.Research Questions:
- How can in-memory computing principles be integrated into SNN architectures to enhance online learning capabilities?
- What are the trade-offs between performance, power, and accuracy when implementing in-memory online learning in SNNs?
- How can the inherent variability and non-ideality of in-memory devices be mitigated or exploited in SNN-based online learning systems?
This research aims to push the boundaries of neuromorphic engineering by combining the strengths of SNNs and in-memory computing. The outcome has the potential to revolutionize ultra-low-power applications, especially in edge devices, wearables, and IoT, making intelligent systems more pervasive and sustainable.