AKANs cut power 50% for analog AI on flexible electronics
New chip design slashes area and power for wearable neural networks
Deep Dive
Analog Kolmogorov-Arnold Networks (AKANs) are introduced for low-power analog function approximation in flexible electronics, using hardware-software co-optimization with circuit-level error modeling and pruning. Validation shows up to 55% area and 50% power savings, with average reductions of nearly 30% across datasets, while pruning can also improve approximation accuracy by regularizing spline parameters.
Key Points
- AKANs achieve up to 55% area and 50% power savings over baseline analog implementations
- The co-optimization approach includes circuit-level error modeling and hardware-software pruning
- Pruning improves accuracy by regularizing spline parameters, not just reducing cost
Why It Matters
Enables battery-friendly, high-performance AI inference directly on flexible wearable sensors for health and IoT