SNS goal

The main objective of the project is to “develop self-healing ultra-low-power Edge-AI systems with neuromorphic architectures, leveraging cross-layer mechanisms, biological principles, and emerging technologies for robust IoT-edge audio and vision applications, reaching a 100X improvement in energy efficiency.”

Workflow

WP1. Demonstrate self-healing ultra-low-power Edge-AI systems by means of cross-layer mechanisms (learning, mapping, circuits, and architectures) applied to advanced node technologies with emerging design styles of analog and in-memory computing. SNNs self-healing mechanisms are specifically tailored on industry IoT-edge audio and vision applications and establish baselines on how drift impacts performance. These developments are characterized with specific KPIs, including graceful degradation of trained SNNs.

WP2. Develop and implement a range of self-healing mechanisms. SNS will integrate a range of biological principles including online learning, local learning rules, context- and attention mechanisms, and homeostatic plasticity, to implement the compensatory changes at different stack layers and ensure stable operation. The resulting algorithms will operate autonomously, without requiring offline retraining. Initial algorithm developments will be done in software, allowing quick identification of promising strategies which are then streamlined for hardware deployment.

WP3. Validate the effectiveness of self-healing at the system and microsystem levels. We will design a new neuromorphic architecture, with CIM blocks, with new mapping strategies exploiting sparsity and data reuse for SNNs. We will also use in-the-loop strategies effectiveness: self-healing algorithms run in software and in hardware, with actual hardware `knobs’ adapted in a closed-loop manner.

WP4. Design micro-architecture level self-healing technologies for ultra-low power neuromorphic computing systems. While SNS develops general approaches applicable to a diversity of emerging materials, we will demonstrate self-healing capabilities on analog using emerging neuromorphic substrate (e.g. in-memory computing with ReRAM- based cells, and ferroelectric models1). To that end, we will build upon quantitative models and hardware data at: 1) Device-level: reflecting the dynamics of plasticity under pulse programming conditions as well as reliability issues; 2) Memory Array level: generating energy and performance metrics. These models will be calibrated with cutting-edge hardware, designed by the SNS consortium.

Target Application Domains