💻 computer science

A2SG:Adaptive and Asymmetric Surrogate Gradients for Training Deep Spiking Neural Networks

The paper proposes A2SG, a unified framework employing adaptive and asymmetric surrogate gradients to address training challenges in deep spiking neural networks by reducing gradient variation and loss landscape curvature, thereby consistently improving accuracy and energy efficiency across diverse models and tasks.

Original authors: Yechan Kang, Yongjin Kweon, Mingyeong Seo, Sohee Park, Yeonguk Jeon, Jongkil Park, Hyun Jae Jang, Jaewook Kim, YeonJoo Jeong, Suyoun Lee, Seongsik Park

Published 2026-06-11

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Yechan Kang, Yongjin Kweon, Mingyeong Seo, Sohee Park, Yeonguk Jeon, Jongkil Park, Hyun Jae Jang, Jaewook Kim, YeonJoo Jeong, Suyoun Lee, Seongsik Park

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

The Big Picture: Teaching a Nervous System to Think

Imagine you are trying to teach a robot brain (a Spiking Neural Network, or SNN) to recognize pictures. Unlike standard computer brains that process information continuously like a flowing river, this robot brain works like a nervous system: it only fires "spikes" (tiny electrical signals) when it really needs to. This makes it incredibly energy-efficient, like a solar-powered watch compared to a high-definition TV.

However, teaching this nervous system is a nightmare. The math used to correct its mistakes (called gradients) is jagged, unstable, and often points in the wrong direction. It's like trying to walk up a mountain where the ground keeps shifting under your feet, and sometimes the path points you straight down a cliff.

The authors of this paper propose a new training method called A2SG (Adaptive and Asymmetric Surrogate Gradients). Think of A2SG as a smart, flexible guide that helps the robot brain find a smooth, stable path to the top of the mountain.

The Two Main Problems

The paper identifies two specific reasons why training these networks is so hard:

The "Jagged Mountain" Problem (Sharp Loss Landscapes):
When standard methods try to train these networks, they often get stuck in "sharp" valleys. Imagine a valley with steep, jagged walls. If you drop a ball there, it bounces wildly and might roll out the other side easily. In machine learning, this means the model is unstable and doesn't generalize well (it memorizes the training data but fails on new data).
- The Cause: The math used to approximate the "spike" is too rigid and creates these sharp, unstable curves.
The "Confused Time Traveler" Problem (Temporal Gradient Confusion):
SNNs process information over time (like a video, not a still photo). The standard training method looks at the whole video at once to figure out mistakes. But the "clues" (gradients) from the beginning of the video often contradict the clues from the end. It's like a detective trying to solve a crime where the witness at 9:00 AM says "The suspect wore a red hat," and the witness at 9:05 AM says "The suspect wore a blue hat." The detective gets confused and can't make a decision.

The Solution: A2SG

The authors introduce a two-part strategy to fix these issues.

1. The Adaptive Guide (Fixing the "Jagged Mountain")

Instead of using a rigid, one-size-fits-all rule for correcting mistakes, A2SG uses an Adaptive approach.

How it works: Imagine the training process is a hiker navigating a foggy forest. The hiker needs to know how wide their "safety zone" (the effective window) should be.
- If the ground is too shaky (high variation in gradients), the guide shrinks the safety zone to focus on stability.
- If the path is clear, it widens the zone to move faster.
The Result: This dynamic adjustment smooths out the "jagged mountain," turning those sharp, dangerous cliffs into gentle, rolling hills. This helps the model settle into a "flat minimum"—a wide, stable valley where the model is robust and doesn't easily get knocked off course.

2. The Asymmetric Guide (Fixing the "Confused Time Traveler")

The second part is Asymmetric. Standard methods treat all neurons the same, regardless of how "excited" they are.

The Metaphor: Imagine a classroom of students. Some are barely awake (low energy), while others are on the verge of shouting out an answer (high energy, close to firing a spike).
- Old Method: The teacher gives the same amount of attention to the sleepy student and the excited student.
- A2SG Method: The teacher realizes the excited student is closer to the answer. They give that student more attention (a larger gradient) because they are more likely to be right. The sleepy student gets less attention.
The Result: By focusing on the neurons that are "ready to fire," the training becomes more efficient. It also aligns the clues from different times (timesteps), so the "detective" no longer gets confused by conflicting witness statements. The path forward becomes clear and consistent.

What Did They Prove?

The paper doesn't just guess; they used math to prove their ideas work:

Smoother Paths: They showed that their "Asymmetric" method mathematically creates less "noise" (variation) than the old "Symmetric" methods. Less noise means a smoother path to the solution.
Flat Minima: They proved that by reducing this noise, the model naturally settles into those "flat valleys" (flat minima) which are known to make AI smarter and more reliable.

The Results: Does It Work?

The authors tested A2SG on many different tasks, from recognizing images (like CIFAR-10) to complex video analysis (neuromorphic datasets) and even dividing images into objects (segmentation).

Better Accuracy: In almost every test, the models trained with A2SG got higher scores than those trained with old methods.
Energy Efficiency: Because the method is smarter, the network fires fewer unnecessary spikes. It's like a car that gets better gas mileage because the driver knows exactly when to accelerate and when to coast.
Versatility: It worked on different types of network architectures, from simple ones (CNNs) to complex ones (Transformers).

Summary

Think of A2SG as a new, smarter way to teach a nervous system. Instead of shouting instructions at the whole brain at once with a rigid, confusing rulebook, it:

Adapts its teaching style based on how shaky the ground is (smoothing out the path).
Prioritizes the neurons that are closest to firing (focusing on the most relevant signals).

The result is a robot brain that learns faster, makes fewer mistakes, and uses less energy to do the job.

Technical Summary: A2SG: Adaptive and Asymmetric Surrogate Gradients for Training Deep Spiking Neural Networks

1. Problem Statement

Training deep Spiking Neural Networks (SNNs) via direct backpropagation (Spatio-Temporal Backpropagation, STBP) faces two primary challenges that hinder performance and generalization compared to traditional Deep Neural Networks (DNNs):

Sharp Loss Landscapes: The use of surrogate gradients to approximate the non-differentiable spiking function often leads to high curvature in the loss landscape. Theoretical analysis in the paper suggests that symmetric surrogate functions with narrow effective windows amplify the Hessian magnitude, driving optimization toward sharp minima which are associated with poor generalization.
Temporal Gradient Confusion: In STBP, parameter updates aggregate gradients across all timesteps. Inconsistent gradient directions across these timesteps create conflicting signals, termed "temporal gradient confusion," which destabilizes optimization and degrades learning performance.
Limitations of Existing Methods: Current approaches to improve surrogate gradients often focus on gradient sparsity or impose high computational overhead (e.g., Dspike). Furthermore, few methods explicitly address the impact of surrogate function shapes on generalization or the temporal consistency of gradients.

2. Methodology: A2SG Framework

The authors propose A2SG (Adaptive and Asymmetric Surrogate Gradients), a unified framework designed to address the above limitations through two core components:

2.1. Spatio-Temporal Adaptive Surrogate Gradients (ST-ASG)

This component dynamically adjusts the effective window width ( $\beta$ ) of the surrogate gradient function to optimize gradient stability.

Spatial Adaptation: At the final timestep ( $T$ ), where activations are relatively stable, the method minimizes Spatial Gradient Variation (SGV). This is defined as the ratio of the variance to the mean of the backpropagated error. Reducing SGV is theoretically linked to reducing the curvature of the loss landscape.
Temporal Adaptation: For preceding timesteps ( $t < T$ ), the method maximizes Temporal Gradient Consistency (TGC), defined as the cosine similarity between gradients at adjacent timesteps. This aligns gradient directions across time to mitigate temporal confusion.
Implementation: The optimal $\beta$ for each layer and timestep is selected using a Bayesian search strategy. This allows the system to adapt to the unpredictable dynamics of training, where the relationship between $\beta$ and gradient metrics varies across layers and epochs.

2.2. Asymmetric Surrogate Gradients (ASY)

The authors introduce an asymmetric surrogate function that accounts for neuronal dynamics, specifically the accumulation of membrane potential.

Design: Unlike symmetric functions (e.g., Boxcar, Triangle) that treat distances from the threshold equally, the ASY function assigns larger gradients to neurons with higher accumulated membrane potentials (closer to firing).
Theoretical Basis: The paper proves that under area and boundary constraints, the asymmetric function yields a lower Coefficient of Variation (CV) for local gradients compared to symmetric counterparts. Lower CV is shown to correlate with flatter loss landscapes.
Mechanism: By concentrating gradients on high-potential neurons, the ASY function reduces unnecessary variability and suppresses gradients for low-activity neurons, promoting sparser and more stable updates.

3. Key Contributions

Theoretical Analysis of Sharpness: The paper establishes a direct mathematical link between local gradient variation and the curvature of the loss landscape. It demonstrates that symmetric surrogates with narrow windows inherently lead to sharper minima ( $O(x^2/\beta^2)$ scaling) compared to DNNs.
Proof of Asymmetry: The authors provide theoretical proofs (Theorems 4.1 and 4.2) showing that asymmetric surrogate gradients achieve lower gradient variation than symmetric ones under specific conditions ( $L\kappa > \sigma^2$ ), thereby promoting convergence to flatter minima.
Unified Framework: A2SG combines spatio-temporal adaptation with an asymmetric design, offering a principled strategy to stabilize optimization and improve generalization without the heavy computational cost of previous adaptive methods.
Comprehensive Evaluation: The method is validated across diverse architectures (CNNs like VGG16, ResNet; Transformers like E-SpikeFormer) and tasks (image classification on CIFAR/ImageNet, neuromorphic datasets, and semantic segmentation).

4. Experimental Results

Extensive experiments demonstrate that A2SG consistently outperforms state-of-the-art methods:

Accuracy Gains: On CIFAR-10 and CIFAR-100, A2SG achieves higher accuracy than existing adaptive surrogate methods (e.g., Dspike, CPNG, LSG). For instance, on ResNet19 with CIFAR-10, A2SG reaches 96.74% accuracy, outperforming LSG (95.17%) and other baselines.
Energy Efficiency: A2SG reduces the total number of spikes generated during inference, indicating improved energy efficiency. For example, on VGG16/CIFAR-10, it reduces spike counts from ~94.6k (baseline) to ~84.9k.
Generalization and Robustness:
- Flat Minima: Analysis of the Fisher Information Matrix (FIM) shows A2SG converges to regions with lower eigenvalues (flatter minima) compared to symmetric surrogates.
- Noise Robustness: Under deletion noise (removing a fraction of spikes), A2SG maintains higher accuracy and exhibits lower weight variance than the baseline, suggesting better generalization.
- Segmentation: On the ADE20K dataset, A2SG improves mean Intersection over Union (mIoU) and reduces power consumption compared to E-SpikeFormer.
Efficiency: The computational overhead of the Bayesian search is amortized over the forward pass, resulting in a relative latency increase of only ~15% at $T=4$ , which decreases as the number of timesteps increases.

5. Significance and Claims

The paper positions A2SG as a general and reliable solution for training deep SNNs. Its significance lies in:

Bridging the Gap: It narrows the performance gap between deep SNNs and DNNs by addressing fundamental issues in surrogate gradient design rather than just tuning hyperparameters.
Principled Optimization: By linking gradient variation to loss landscape curvature, the work provides a theoretical foundation for why asymmetric and adaptive designs improve generalization.
Scalability: The method is shown to be compatible with various neuron models (LIF, PLIF) and training strategies (tdBN, RMP-loss), making it a versatile tool for the broader neuromorphic computing community.

The authors conclude that while A2SG significantly improves training stability, accuracy, and efficiency, formal convergence proofs for the specific error bounds of the asymmetric surrogate remain an open problem for future research.