⚡ electrical engineering

Neuromorphic Energy-Aware Learning for Adaptive Deep Brain Stimulation

This paper introduces an energy-aware learning framework that co-optimizes actuator stimulation and inference efficiency by incorporating actuator energy into reinforcement learning rewards and deploying a compressed spiking neural network on neuromorphic hardware, achieving significant reductions in both pathological oscillations and power consumption for adaptive deep brain stimulation in Parkinson's disease.

Original authors: Binh Nguyen, Colleen Josephson, Mircea Teodorescu, Gert Cauwenberghs, Jason Eshraghian

Published 2026-06-30

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Binh Nguyen, Colleen Josephson, Mircea Teodorescu, Gert Cauwenberghs, Jason Eshraghian

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, battery-powered robot living inside a person's brain. Its job is to act as a "traffic cop" for the brain's electrical signals. When the brain gets stuck in a chaotic rhythm (which happens in Parkinson's disease), this robot needs to send out gentle electrical pulses to smooth things out. This is called Deep Brain Stimulation (DBS).

For years, scientists have been trying to make the robot's "brain" (the computer chip) smarter and more energy-efficient. But there's a catch: the robot's arm (the part that sends the pulses) uses way more energy than its brain does.

Think of it like a hybrid car. You can build the most fuel-efficient engine in the world (the controller), but if the car is constantly dragging a heavy anchor behind it (the stimulation), you're still going to run out of gas. The paper argues that to save battery life, you can't just make the brain efficient; you have to teach the brain to be lazy about how it uses the anchor.

Here is how the researchers solved this problem, broken down into simple steps:

1. The "Energy-Aware" Teacher

The researchers created a new way to train the robot's brain. Usually, you train a robot by saying, "Good job if you fix the tremor!" But this new method adds a second rule: "Good job if you fix the tremor AND you didn't use much electricity to do it."

They call this Energy-Aware Learning. It's like training a delivery driver not just to get the package to the house, but to do it while driving the most fuel-efficient route possible. If the driver takes a shortcut that saves gas but misses the house, they get a bad score. If they drive a huge, gas-guzzling truck to the house, they also get a bad score. They have to find the perfect balance.

2. The "Spiking" Brain vs. The "Standard" Brain

To make the robot small enough to fit in a human body, they used a special type of computer chip called Neuromorphic hardware.

Standard Chips (like in your phone): These are like a lightbulb that stays on 24/7, even when there's nothing to look at. They constantly check the brain's signals, even when the brain is quiet. This wastes a lot of battery.
Neuromorphic Chips (The Xylo): These are like a motion-sensor light. They only "wake up" and do work when a specific signal (a "spike") happens. If the brain is quiet, the chip sleeps. This is incredibly efficient.

The researchers trained a "Spiking Neural Network" (SNN) on this motion-sensor chip. It learns to listen to the brain's natural, chaotic rhythm and only send a pulse when absolutely necessary.

3. The "Distillation" Trick (Teaching the Student)

The "teacher" robot was smart but a bit too big and chatty for the tiny implant. So, the researchers used a trick called Knowledge Distillation.
Imagine a brilliant professor (the Teacher) who knows everything but talks too much. They hire a student (the Student) to learn the material. The goal isn't just for the student to get the right answers, but to learn how to get those answers using fewer words.
They forced the student to be "sparse"—meaning it had to fire very few electrical signals to do the job. The result was a tiny, super-efficient student robot that could fit on the tiny chip without draining the battery.

4. The Results: A Miracle of Efficiency

When they tested this new system in a computer simulation of a Parkinson's brain, the results were impressive:

Stopping the Chaos: The robot successfully reduced the bad brain rhythms by 45%, which is a huge improvement for the patient.
Saving the Battery: Because the robot learned to be lazy, it reduced the amount of electricity sent to the brain by 80% compared to old, "always-on" methods.
The Chip's Power: The tiny chip used so little power that if you put it in a standard implant battery, it could run for over a year without needing a surgery to replace the battery. In contrast, if they tried to run this on a standard computer chip (like the one in a high-end smartphone), the battery would die in less than an hour.

The Big Picture

The paper's main message is that you can't just make the computer brain efficient; you have to make the whole system efficient. By teaching the AI to care about the cost of the electricity it sends out, they created a system that is both a great doctor (fixing the disease) and a great accountant (saving the battery).

This isn't just about making a better chip; it's about realizing that in a medical implant, the "actuator" (the part that does the work) is often the biggest energy hog. If you don't teach the AI to be gentle with that part, all the other efficiency gains don't matter. This new approach solves both problems at once.

Technical Summary: Neuromorphic Energy-Aware Learning for Adaptive Deep Brain Stimulation

Problem Statement

The deployment of neural network controllers on edge and neuromorphic architectures has successfully reduced computational inference costs. However, in physical closed-loop systems like Deep Brain Stimulation (DBS) for Parkinson's Disease (PD), the energy expended by the actuator (the neurostimulator) often rivals or exceeds the energy required for inference once the controller runs on efficient hardware. Conventional DBS devices deliver continuous high-frequency pulses regardless of clinical state, leading to battery wastage and stimulation-induced side effects. While adaptive DBS (aDBS) modulates parameters based on biomarkers, existing approaches often rely on conventional hardware and window-aggregated processing that discards fine-grained temporal spike structure. Furthermore, reducing inference energy alone is insufficient; an effective aDBS controller must co-optimize therapeutic efficacy with the physical charge delivered by the stimulator, as the actuator becomes the dominant power demand in implantable systems.

Methodology

Biophysical Simulation Environment

The authors validated control policies using a biophysically detailed simulation of the rat Cortico-Basal Ganglia-Thalamocortical (CBGT) circuit, based on the 6-hydroxydopamine (6-OHDA) lesioned rat model. The network comprises 80 single-compartment neurons across eight populations (cortical, STN, GPe, GPi, Striatum, Thalamus) modeled with Hodgkin-Huxley dynamics. Parkinsonian states were induced by reducing striatal M-current and corticostriatal strength while increasing intra-GPe inhibition, reproducing exaggerated $\alpha$ - $\beta$ (7–35 Hz) oscillations.

Deep Spiking Q-Network (DSQN) Architecture

A Deep Spiking Q-Network (DSQN) was implemented to select discrete stimulation parameters (frequency, pulse width, amplitude).

Input: The network receives an 80-channel binary spike matrix from the CBGT model, spatially pooled to 16 channels via average pooling to match hardware constraints.
Architecture: Two hidden layers of 128 Leaky Integrate-and-Fire (LIF) neurons with explicit synaptic current dynamics.
Action Decoding: A partitioned population coding scheme decodes actions. The output layer (9 neurons) is split into three sub-populations corresponding to frequency, pulse width, and amplitude. A winner-take-all (WTA) readout over the simulation window determines the discrete action (decrease, maintain, increase).
Training: The agent uses off-policy Q-learning with an experience replay buffer, optimized via surrogate gradients in PyTorch/Rockpool.

Energy-Aware Reward Formulation

The core innovation is a reward function that jointly optimizes symptom suppression and energy conservation. The reward $r_t$ is discontinuous:

If the instantaneous $\beta$ -band power ( $\beta_t$ ) exceeds a threshold ( $\tau_\beta$ ), the agent receives a penalty proportional to the excess.
If $\beta_t \le \tau_\beta$ , the agent receives a reward based on therapeutic efficacy and a normalized energy term ( $E_t/E_{max}$ ), where $E_t$ is the Root Mean Square (RMS) of the stimulation current. This forces the policy to minimize charge delivery only after ensuring therapeutic suppression.

Sparsity-Constrained Knowledge Distillation

To deploy the policy on ultra-low-power hardware, a dense "teacher" DSQN was distilled into sparse "student" SNNs.

Loss Function: The total loss combines Kullback-Leibler (KL) divergence between teacher and student Q-values and a sparsity penalty.
Sparsity Penalty: A Bernoulli KL divergence term regularizes the average firing rate of hidden layers toward a target sparsity $\rho$ , minimizing Synaptic Operations (SynOps).
Hardware Mapping: The distilled student networks were quantized (8-bit precision) and deployed on the SynSense XyloAudio 3 neuromorphic processor.

Benchmarking

The neuromorphic implementation was benchmarked against:

Conventional DBS (cDBS): Continuous stimulation at fixed parameters.
Clinical aDBS: Dual-threshold heuristic toggling stimulation based on $\beta$ -power.
Artificial Neural Network (ANN): Rate-coded feedforward DQN.
Recurrent Neural Network (RNN): GRU-based DQN.
All were compared on an NVIDIA Jetson Orin Nano (CUDA) for power and latency metrics.

Key Contributions

Energy-Aware Control: The framework introduces a reinforcement learning reward that directly penalizes stimulation charge delivery. This ensures the policy optimizes the first-order power demand of the implant (the actuator) rather than just the inference cost.
Neuromorphic Benchmarking: The study demonstrates a 28.1 $\times$ reduction in energy per inference when deploying the SNN on the SynSense XyloAudio 3 compared to an equivalent ANN on the Jetson Orin Nano, while operating within the sub-milliwatt power envelope required for chronic implants.
Biophysical Validation: The system achieves a 45.2% reduction in pathological $\alpha$ - $\beta$ band power and an 80.0% reduction in cumulative stimulation charge relative to continuous DBS in a biophysical CBGT model.

Results

Therapeutic Efficacy

In the dopamine-depleted biophysical model, the SNN controller reduced integrated $\alpha$ - $\beta$ band power by 45.2% (95% CI [42.6, 47.7]) across 10 seeds. The controller rapidly detected pathological states and initiated high-frequency stimulation to break recurrent synchrony between the STN and GPe. Unlike the unstimulated baseline, the SNN restored a desynchronized firing regime.

Stimulation Energy Efficiency

Charge Reduction: The SNN achieved an 80.0% reduction in cumulative charge delivery compared to cDBS.
Comparison to Heuristics: The SNN outperformed a clinical dual-threshold aDBS heuristic, delivering 53.5% less charge while maintaining deeper $\beta$ -suppression (85.9% reduction vs. 56.2% for the heuristic).
Baseline Failures: The ANN baseline converged to a degenerate strategy (minimal stimulation, no $\beta$ suppression), while the RNN baseline adopted a static, state-agnostic strategy (maximum parameters continuously), consuming 50.6% more energy than cDBS.

Hardware Efficiency and Compression

Sparsity Trade-off: Through sparsity-constrained distillation, the best student policy achieved equivalent therapeutic efficacy (45.9% $\beta$ -reduction) to the teacher at 16.8 SynOps/ms, a 3.6 $\times$ reduction in synaptic throughput.
Power Consumption: The Xylo SNN drew 0.52 mW of mean system power, compared to ~5.7 W for the Jetson baselines.
Energy per Inference: The neuromorphic implementation consumed 1.64 $\times$ 10 $^{-4}$ J per inference, representing a 28.1 $\times$ reduction compared to the ANN and 72.5 $\times$ compared to the RNN.
Battery Life Projection: With the SNN policy, the total system draw (stimulation + inference) is projected to be ~0.57 mW, enabling over one year of continuous operation on a standard 5 Wh non-rechargeable IPG battery. In contrast, Jetson-based controllers would deplete the battery in under an hour.

Significance and Claims

The paper claims that hardware efficiency is a learnable property of the neural controller. By incorporating actuator energy directly into the learning objective, the framework addresses both major power demands in implantable neuromodulation: inference and stimulation.

The authors assert that reducing inference cost alone is necessary but not sufficient; once the controller is efficient, the actuator becomes the cost worth reducing. The proposed co-optimization framework is distinct from approaches that optimize stimulation protocols or inference efficiency in isolation. The results demonstrate that a neuromorphic controller can operate within strict thermal and battery constraints (sub-milliwatt envelope) while discovering charge-efficient strategies that standard objectives would miss. The work suggests that for chronic implants, the latency trade-off of neuromorphic hardware is acceptable given the critical constraints of power and energy, resolving the hardware barrier that previously precluded neural network controllers from such applications. The energy-aware reward formulation is presented as transferable to other responsive neurostimulation systems (e.g., epilepsy, pain) where actuator energy becomes the dominant cost.