TLDR: ASPEN is a novel technique for neuromorphic systems that enables ultra-low-power, always-on operation in wearables. It achieves this by adaptively adjusting neuron firing thresholds during inference, a method inspired by biological intrinsic plasticity. This approach significantly reduces energy consumption and spike counts while maintaining high accuracy, offering superior robustness and extended battery life compared to traditional fixed-threshold systems.
Wearable technology has become an integral part of our daily lives, from fitness trackers to health monitors, with hundreds of millions of units shipped annually. These devices offer continuous, real-time monitoring of various physiological signals, but their ability to perform complex AI tasks while operating for weeks or months on tiny batteries presents a significant challenge. Traditional computing methods, even optimized ones, struggle to meet the ultra-low-power demands of always-on sensing.
A common workaround, offloading data to the cloud, introduces issues like high latency, significant wireless energy consumption, and privacy concerns. While conventional edge computing attempts to localize inference, it still relies on continuous, clock-driven computation, which isn’t ideal for extreme energy constraints. Existing optimization techniques like network pruning or quantization are static and don’t adapt to fluctuating energy availability.
This is where neuromorphic systems offer a promising alternative. Inspired by the human brain, these systems operate on an event-driven model, communicating through discrete electrical pulses called “spikes.” Computation only activates in response to informative input events, allowing for deep idle states and minimal baseline power consumption. This paradigm is particularly well-suited for low-power, always-on tasks in constrained environments, with platforms like Intel’s Loihi 2 and SynSense’s Xylo-IMU demonstrating significant energy reductions compared to traditional CPUs or GPUs.
Introducing ASPEN: Adaptive Spiking for Energy-Aware Neuromorphic Systems
Despite their potential, most current Spiking Neural Networks (SNNs) in neuromorphic systems use static neuron configurations, with fixed parameters like firing thresholds. These parameters are typically set during training and cannot adapt dynamically to changing energy availability or application demands. This static nature prevents neuromorphic systems from fully realizing their potential for adaptive, energy-aware computing in real-world scenarios.
Addressing this limitation, researchers from the University of Massachusetts Amherst – Eduardo Calle-Ortiz, Hui Guan, Deepak Ganesan, and Phuc Nguyen – have introduced ASPEN. This novel technique is designed to enable always-on inference in wearable systems operating at microwatt-level power. ASPEN draws inspiration from biological observations of “intrinsic plasticity,” where neurons adjust their excitability to maintain balance and respond to varying stimuli. Unlike “extrinsic plasticity,” which involves changes in synaptic weights (the strength of connections between neurons), intrinsic plasticity focuses on modulating internal properties like firing thresholds.
How ASPEN Works: Learning to Adapt
ASPEN’s core innovation lies in its approach to training and deployment. During training, it introduces “stochastic perturbations” – essentially, random variations – to neuron firing thresholds. This process encourages the network to learn robust internal representations that can perform well across a range of spiking conditions. By exposing the network to diverse excitability levels during training, it becomes less sensitive to specific threshold values, making it more adaptable.
At deployment time, ASPEN allows real-time control over energy usage by dynamically adjusting spiking activity without needing to retrain the model or modify its architecture. Since each spike incurs computational and power costs, minimizing the number of spikes is crucial for operating under constrained energy budgets. ASPEN’s adaptive threshold adjustment directly regulates spike generation, allowing it to balance energy consumption and classification accuracy.
The framework supports two complementary energy-aware adaptation strategies:
- Single-Model Adaptation: A network trained with stochastic thresholds can adjust its internal thresholds post-deployment to match current energy constraints. This provides a continuous and graceful trade-off between accuracy and spiking activity.
- Model-Switching Adaptation: For broader adaptability, ASPEN can select among multiple pre-trained models, each optimized for a specific spiking regime. This allows the system to transition between models that offer better accuracy-spike trade-offs for different operating points, effectively creating a “Pareto front” of optimal performance across various energy budgets.
Real-World Impact and Validation
The researchers evaluated ASPEN on neuromorphic emulators and hardware, demonstrating significant reductions in spike counts and energy consumption while maintaining accuracy comparable to state-of-the-art methods. For instance, ASPEN reduced energy consumption to approximately 120 microwatts (µW). This is about 1000 times lower power than traditional von Neumann-based embedded systems and twice as low as other state-of-the-art neuromorphic systems like SynSense-IMU.
These energy savings translate directly into extended battery life for wearable devices. A standard 100 mAh battery, commonly found in devices like the Xiaomi Mi Band 3 or Samsung Galaxy Fit, could power an ASPEN-enabled IMU for approximately 128 days (over four months) of continuous operation. Even a smaller 22 mAh battery, like that in the Oura Ring, could support continuous activity recognition for up to 22 days. This is a significant improvement, as many commercial devices operate intermittently to conserve power rather than continuously.
Cross-dataset validation on various human activity recognition datasets (KU-HAR, UCI-HAR, UniMiB-SHAR ADL) further confirmed ASPEN’s generalizability and robustness. The models consistently outperformed fixed-threshold baselines, especially when operating away from the initial training threshold. Discrete threshold sampling strategies, in particular, proved highly efficient and compatible with hardware, offering optimal accuracy-spike operating points across different energy regimes.
Also Read:
- Advancing Spiking Neural Networks with Single-Timestep Processing and Adaptive Optimization
- XR-NPE: Enhancing Extended Reality Performance with Advanced Chip Design
Conclusion
ASPEN represents a significant step forward in making intelligent, always-on, ultra-low-power wearables a reality. By leveraging bio-inspired intrinsic plasticity and adaptive threshold modulation, it provides a lightweight and scalable technique for dynamic energy control in neuromorphic systems. This adaptability, combined with remarkable energy efficiency, paves the way for a new generation of maintenance-free wearable devices capable of continuous, reliable operation in diverse real-world environments.
