Enabling Green Wireless Communications with Neuromorphic Continual Learning

TL;DR

SpikACom (Spiking Adaptive Communications), a neuromorphic computing framework that synergizes brain-inspired spiking neural networks (SNNs) with wireless signal processing to deliver sustainable intelligence, is introduced, providing evidence that neuromorphic computing can empower the sustainability of modern digital systems.

Abstract

The pursuit of carbon-neutral wireless networks is increasingly constrained by the escalating energy demands of deep learning-based signal processing. Here, we introduce SpikACom (Spiking Adaptive Communications), a neuromorphic computing framework that synergizes brain-inspired spiking neural networks (SNNs) with wireless signal processing to deliver sustainable intelligence. SpikACom advances the paradigm shift from energy-intensive, continuous-valued processing to event-driven sparse computation. Moreover, it supports continual learning in dynamic wireless environments via a dual-scale mechanism that integrates channel distribution-aware context modulation with a synaptic consolidation rule using SNN-specific statistics, mitigating catastrophic forgetting. Evaluations across critical wireless communication tasks, including semantic communication, multiple-input multiple-output (MIMO) beamforming, and channel estimation demonstrate that SpikACom matches full-precision deep learning baselines while achieving an order-of-magnitude improvement in computational energy efficiency. Our results position SNNs as a promising pathway toward green wireless intelligence, providing evidence that neuromorphic computing can empower the sustainability of modern digital systems.

Figures (12)

• Figure 1: Illustration of neuromorphic computing empowered wireless systems. (a) Evolution of communication paradigms. Left: Traditional communication focuses on bit-level reliability and is built upon separate model-driven signal processing blocks, where information is mapped to constellation symbols for transmission. Middle: ANN-based communication employs dense neural networks to replace or augment conventional processing blocks, enabling information processing at higher levels of abstraction. The transmitted representations are intermediate continuous-valued activations. Right: SNN-based (neuromorphic) communication shifts toward event-driven processing, where information is encoded and conveyed as sparse binary spikes, enabling energy-efficient operation and facilitating semantics-oriented, cognition-inspired communication. (b) Comparison between an artificial neuron and a spiking neuron. The spiking neuron communicates via discrete binary spikes and inherently operates in the spatio-temporal domain, whereas the artificial neuron computes with continuous-valued activations. (c) Neuromorphic computing enabled wireless communication system illustrating low-power operation and continual adaptation in dynamic wireless environments affected by factors such as mobility, weather, and environmental scattering.
• Figure 2: Overview of the SpikACom framework. Left: Pilots collected in different wireless conditions are used to identify the environment. Middle: A hypernet-based context modulator senses channel distribution shifts (quantified via earth mover's distance (EMD)) based on the pilots and generates binary gates to modulate the backbone SNN. Right: The backbone SNN processes the task inputs and supports task-dependent incorporation of domain knowledge to enhance performance, while spiking rate consolidation (SRC) regularizes synaptic weights based on neuronal activity to mitigate catastrophic forgetting.
• Figure 3: Performance evaluation of SpikACom on neuromorphic semantic communication.(a) System overview: Event data collected by distributed dynamic vision sensor (DVS) cameras is encoded by SNNs and transmitted over varying wireless channels with different power-delay profiles (PDPs) to the access point (AP). (b)--(d) Evolution of test accuracy versus the number of learned environments. Each curve tracks the test accuracy of a specific environment $i$ evaluated after the model has learned up to environment $k$ (where $k \ge i$). Comparisons include (b) Vanilla fine-tuning, (c) EWC, (d) SpikACom (ours). The environments are arranged in descending SNR, ranging from $8$ dB (Env 1) to $-20$ dB (Env 8). (e) Ablation study of SpikACom. The y-axis represents the average accuracy across all environments after the sequential learning process is completed. (f) Performance comparison of SNN against ANN baselines (Bi-RNN, Bi-LSTM, Transformer) and traditional LDPC coding under different SNRs. (g) Inference energy comparison. The energy consumption of SNN exhibits negligible scaling with the network size compared with ANNs.
• Figure 4: Performance evaluation of SpikACom on multi-user MIMO beamforming.(a) System overview: A base station (BS) forms beams serving multiple users located in different wireless environments. The complex-valued channel state information is encoded into spikes for energy-efficient processing, and a model-driven module is integrated to improve beamforming performance. (b)--(d) Sum-rate evolution over sequential environments. Comparisons include (b) Vanilla fine-tuning, (c) EWC, and (d) SpikACom (ours). The channels are generated based on the 3D ray-tracing DeepMIMO dataset, where each environment corresponds to a distinct user location distribution. (e) Ablation study. The y-axis shows the average sum rate across all environments after the sequential learning process is completed. (f) Sum-rate performance versus the maximum transmit power $P_{T}$. We compare the proposed SNN + domain knowledge (DK) against ANN + DK, SNN without DK, and traditional beamforming benchmarks (WMMSE, RBD, RZF, and MMSE). (g) Energy consumption comparison of analyzed schemes for beamforming. The energy consumption of SNN+DK is reduced by over $10\times$ compared with the ANN benchmark and is comparable to low-complexity traditional baselines (e.g. MMSE).
• Figure 5: Performance evaluation of SpikACom on channel estimation in OFDM systems.(a) System overview: Sparse pilots are inserted into the OFDM frequency-time grid and transmitted through dynamic wireless channels. The received pilots are processed by the proposed SNNResNet, which uses spike-based residual convolutional blocks for feature extraction and incorporates an end-to-end LMMSE-inspired residual connection to facilitate channel reconstruction. (b)--(d) MSE evolution over sequential environments. Comparisons include (b) Vanilla fine-tuning, (c) EWC, and (d) SpikACom (ours). The WINNER II channel model is adopted, where each environment corresponds to a distinct propagation scenario. Lower MSE indicates better estimation performance. (e) Ablation study. The y-axis represents the average MSE, computed using the arithmetic mean, across all environments after the sequential learning process is completed. (f) Performance comparison among SNN-based channel estimation (SNNResNet), ANN-based channel estimation (ChannelNet and ReEsNet), and traditional estimators (LS and LMMSE). (g) Energy consumption comparison. SNN-based scheme is $7\times$ more energy-efficient than lightweight ANN baseline (ReEsNet).

Theorems & Definitions (5)

• Lemma 1
• Theorem 1
• Remark 1
• Lemma 2
• Theorem 2