⚡ electrical engineering

Energy-Aware Multi-Exit TinyML for Smart Zero-Energy Devices

This paper presents an energy-aware Multi-Exit TinyML framework for zero-energy devices that dynamically scales computational effort based on input difficulty and regulates system operation via auxiliary circuits, achieving a 29.6% reduction in energy consumption while maintaining detection accuracy.

Original authors: Shahab Jahanbazi, Mateen Ashraf, Lieven De Strycker, Jeroen Famaey, Onel L. A. Lopez

Published 2026-03-10

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Shahab Jahanbazi, Mateen Ashraf, Lieven De Strycker, Jeroen Famaey, Onel L. A. Lopez

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you have a tiny, solar-powered robot that lives in your house. Its job is to look around and tell you if a person is in the room. But there's a catch: this robot has no battery. It runs entirely on the sunlight (or indoor light) it catches in real-time. If the light dims, the robot might run out of juice and shut down mid-thought.

This paper presents a clever solution to make this robot smarter, more efficient, and less likely to "pass out" from lack of energy. They call it Energy-Aware Multi-Exit TinyML.

Here is the breakdown of how it works, using simple analogies:

1. The Problem: The "All-or-Nothing" Brain

Usually, when a computer looks at a picture, it runs the entire brain process for every single image, no matter how easy or hard the picture is.

The Analogy: Imagine you are a security guard. Someone walks by wearing a bright red shirt. It's obvious they are a person. But your rule is: "I must run a full 10-mile marathon to check every single person, even the ones in red shirts."
The Result: You waste a ton of energy on easy tasks. If the sun goes behind a cloud while you are running that marathon, you might collapse before you finish.

2. The Solution: The "Multi-Exit" Shortcut

The authors built a special brain (a TinyML model) that has two exits.

The Analogy: Think of this brain like a security checkpoint with two gates.
- Gate 1 (The Easy Exit): If the robot sees something very obvious (like a person in a red shirt), it stops right there, says "Yes, that's a person," and goes home. It saves a massive amount of energy.
- Gate 2 (The Deep Dive): If the image is blurry or confusing (like a shadow that might be a person), the robot doesn't stop. It walks through to the second, deeper gate to think harder and get a better answer.
The Benefit: The robot only does the hard work when it absolutely has to. For easy jobs, it takes the shortcut. This saves energy for the "rainy days" when the light is low.

3. The Safety Net: The "Fuel Gauge"

Since the robot runs on harvested light, it needs to know exactly how much "fuel" (energy) is in its tank before it starts a task.

The Analogy: Imagine you are driving a car with no gas tank, just a small cup of gas you fill from a rain barrel. Before you start driving, you check the cup.
- Old Way: You just guess. If you guess wrong, you run out of gas in the middle of the road and the engine dies.
- This Paper's Way: The robot has a super-accurate fuel gauge. Before it starts looking at a picture, it checks: "Do I have enough energy to finish the whole job?"
- The "Double Check": If the robot takes the "Easy Exit" but isn't 100% sure, it pauses and checks the fuel gauge again before deciding to go to the "Deep Dive" exit. This ensures it never starts a long task it can't finish.

4. The Hardware Hack: The "Smart Light Switch"

The robot also needed to stop wasting energy on parts of itself it wasn't using.

The Analogy: Imagine you have a camera that is always on, even when it's sleeping. That's like leaving your house lights on 24/7.
The Fix: The team built a special electronic switch (using a MOSFET) that acts like a master light switch. When the robot isn't taking a picture, it completely cuts the power to the camera. It's like flipping the breaker switch off. This saves so much energy that the robot can run longer on the same amount of sunlight.

5. The Results: Doing More with Less

When they tested this system:

Energy Savings: It used about 30% less energy than previous methods.
More Tasks: Because it saved energy, the robot could take and analyze more photos before running out of power.
Accuracy: It didn't sacrifice accuracy. It was just as good at spotting people, but it was much smarter about how it did it.

Summary

In short, this paper teaches a tiny, solar-powered robot how to be a smart energy manager. Instead of running a marathon for every task, it learns to take a shortcut when the task is easy, checks its fuel tank before starting, and turns off its lights when it's sleeping. This allows it to stay alive and useful even when the sun isn't shining brightly.

1. Problem Statement

The proliferation of smart, autonomous systems has driven the need for Zero-Energy Devices (ZEDs) that operate solely on harvested ambient energy (e.g., solar) without batteries. While these devices offer sustainability, they face severe constraints:

Intermittent Power: Energy availability fluctuates, leading to frequent "brownouts" (power failures) and loss of state.
Resource Scarcity: ZEDs have extremely limited memory (RAM/Flash) and energy budgets (millijoules).
The Accuracy-Energy-Memory Trade-off: Traditional TinyML approaches often require running a full, fixed-depth neural network for every input. This wastes energy on "easy" inputs that could be classified with high confidence using fewer computations. Conversely, running multiple models to adapt to energy levels (as done in some prior works) exceeds the tight memory constraints of ZEDs.

The core challenge is to design a system that maximizes the number of completed tasks (inferences) per unit of harvested energy while maintaining high detection accuracy, all within a single, memory-constrained model.

2. Methodology

The authors propose a unified framework combining Multi-Exit TinyML, Energy-Aware Control Policies, and Hardware Optimization.

A. Multi-Exit TinyML Architecture

Instead of storing multiple models or running a fixed-depth network, the system uses a single model with multiple exit points (inspired by BranchyNet).

Base Model: A MobileNet V1 backbone trained on the Visual Wake Words (VWW) dataset for person detection.
Exit Points: Two exits are implemented:
- EX1 (Early Exit): A lightweight classifier attached after an early convolutional layer. It consumes less energy ( $e_1$ ) but has lower accuracy ( $Acc_1$ ).
- EX2 (Deep Exit): The full network depth. It consumes more energy ( $e_2$ ) but offers higher accuracy ( $Acc_2$ ).
Training: The model is trained using a weighted sum of losses from both exits to ensure the backbone learns features useful for both early and late classification.

B. Energy-Aware Runtime Policy

The system employs a dynamic decision policy at runtime to select the exit point based on input difficulty and available energy:

Confidence Gating: The system first runs inference at EX1.
- If the confidence score ( $O_1$ ) is high (above a threshold $\gamma_2$ ) or very low (below $\gamma_1$ ), the result is accepted immediately.
- If the score is ambiguous (between $\gamma_1$ and $\gamma_2$ ), the system considers escalating to EX2.
Energy Safety Check: Before escalating to EX2, the device performs a second energy measurement.
- If the remaining energy is sufficient to cover the cost of the remaining layers ( $e_2 - e_1$ ), it proceeds to EX2.
- If energy is insufficient, it falls back to the EX1 prediction to prevent a power failure.
  This policy ensures that easy inputs are processed cheaply, while hard inputs receive more computational resources only when energy permits.

C. Hardware and Circuit Design

To support this logic on a severely constrained platform (Arduino Nano 33 BLE Sense with nRF52840):

Power Gating: A custom P-MOSFET-based power-gating circuit is designed to completely disconnect the camera module from the power rail when not in use. This is more energy-efficient than the load-switch approaches used in prior works, as it eliminates leakage current.
Energy Monitoring: The system uses the MCU's ADC to monitor the voltage across the storage capacitor. A resistive divider scales the voltage to safe levels, and a MOSFET switches the divider on only during measurement to minimize leakage.
Execution Scheduling: The system uses an "Adaptive Start Time" rule, allowing the device to defer execution within a time window until sufficient energy has been harvested, rather than forcing execution at a fixed time.

3. Key Contributions

First Energy-Aware Multi-Exit TinyML on ZEDs: The paper presents the first implementation of a multi-exit model specifically optimized for the strict memory and energy constraints of battery-less devices.
Memory-Efficient Design: By using a single model with lightweight auxiliary layers for early exits, the approach avoids the memory overhead of storing multiple models (a common issue in adaptive inference).
Novel Hardware Circuit: The design of a low-leakage, MOSFET-based power-gating circuit for the camera module, which reduces energy consumption by ~34% compared to standard load-switch methods.
Robust Control Policy: A "confidence-gated" policy that dynamically balances inference depth with real-time energy availability, guaranteeing no power failures during execution.

4. Results

The system was prototyped and tested under realistic solar harvesting conditions:

Energy Efficiency: Compared to a state-of-the-art single-exit TinyML approach (Work [2]), the proposed multi-exit method achieved an energy consumption reduction of approximately 29.6% while maintaining comparable accuracy.
Task Completion: Under fixed harvesting conditions, the proposed approach completed a significantly higher number of inference pipelines before the capacitor depleted compared to baseline methods.
Hardware Savings: The custom MOSFET power-gating circuit reduced the energy consumption of the image capture/pre-processing phase by 34% compared to previous load-switch designs.
Accuracy vs. Energy Trade-off: By tuning the ambiguity thresholds ( $\gamma_1, \gamma_2$ ), the system can be tuned to prioritize energy savings (more early exits) or accuracy (more deep exits). The optimal configuration ( $\gamma_1=0.3, \gamma_2=0.7$ ) provided the best balance.

5. Significance

This work bridges the gap between advanced machine learning architectures and the physical realities of zero-energy hardware. It demonstrates that adaptive inference is not just a software optimization but a critical requirement for the viability of smart ZEDs. By jointly optimizing the model architecture, runtime control policy, and hardware power management, the authors prove that accurate, autonomous person detection is feasible on devices that rely entirely on intermittent ambient energy, paving the way for truly maintenance-free, sustainable IoT deployments.