Research map / Continual learning / Class-incremental
ALADE-SNN: Adaptive Logit Alignment in Dynamically Expandable Spiking Neural Networks for Class Incremental Learning
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes ALADE-SNN, a dynamically expandable spiking neural network for class-incremental learning that adds adaptive logit alignment, OtoN weight suppression, and knowledge distillation on top of DER.
Problem
Class-incremental learning in spiking neural networks suffers from catastrophic forgetting and logit bias between old and new tasks caused by limited exemplar memory and fixed features from frozen past extractors.
Approach
Builds a DER-style dynamically expanding SNN and adds an adaptive correction term that re-balances ground-truth logits between old and new classes during classifier fine-tuning, plus OtoN suppression that zeroes classifier weights from frozen old features to new classes during feature-extractor training and releases them during fine-tuning.
Key results
- ALADE-SNN reaches 75.42 ± 0.74% Avg on CIFAR100-B0 with 10 steps, versus 60.47% for DSD-SNN and 75.36% for DER (ANN).
- On CIFAR100-B0 10 steps, ablation shows adaptive logit alignment adds about 1.0 point Avg (74.51 without vs 75.42 with) and OtoN about 0.4 points.
- On DVS-CIFAR10 (2 tasks), ALADE-SNN gets 83.5% Avg and 77.2% Last versus 81.4% and 74.8% for DER on ResNet18.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add adaptive logit alignment (ablation) | ALADE-SNN · CIFAR100-B0-10steps | Avg accuracy (%): $74.51_{\pm 0.76}$ -> $75.42_{\pm 0.74}$ | Helped |
| add adaptive logit alignment (ablation) | ALADE-SNN · CIFAR100-B0-10steps | Last accuracy (%): $61.80_{\pm 0.78}$ -> $63.13_{\pm 0.4}$ | Helped |
| add OtoN suppression (ablation) | ALADE-SNN · CIFAR100-B0-10steps | Avg accuracy (%): $75.30_{\pm 1.20}$ -> $75.42_{\pm 0.74}$ | No clear effect |
| add OtoN suppression (ablation) | ALADE-SNN · CIFAR100-B0-10steps | Last accuracy (%): $62.30_{\pm 0.94}$ -> $63.13_{\pm 0.4}$ | Helped |
| add knowledge distillation (ablation) | ALADE-SNN · CIFAR100-B0-10steps | Avg accuracy (%): $75.51_{\pm 1.31}$ -> $75.42_{\pm 0.74}$ | No clear effect |
| add knowledge distillation (ablation) | ALADE-SNN · CIFAR100-B0-10steps | Last accuracy (%): $63.60_{\pm 0.75}$ -> $63.13_{\pm 0.4}$ | No clear effect |
| add adaptive logit alignment (ablation) | DER-SNN* · CIFAR100-B0-10steps | Avg accuracy (%): $74.22_{\pm1.06}$ -> $75.25_{\pm 1.09}$ | Helped |
| add OtoN suppression (ablation) | DER-SNN* · CIFAR100-B0-10steps | Avg accuracy (%): $74.22_{\pm1.06}$ -> $74.65_{\pm 1.44}$ | Helped |
| add knowledge distillation (ablation) | DER-SNN* · CIFAR100-B0-10steps | Avg accuracy (%): $74.22_{\pm1.06}$ -> $74.47_{\pm 1.85}$ | Helped |
| use ALADE-SNN instead of DSD-SNN | CIFAR100-B0-10steps · CIFAR100-B0-10steps | Avg accuracy (%): 60.47% to 75.42% (+14.95%) | Helped |
| use ALADE-SNN instead of DER | CIFAR100-B50-10steps · CIFAR100-B50-10steps | Avg accuracy (%): 72.81 -> 71.50 | Hurt |
| use ALADE-SNN instead of DER | CIFAR100-B50-5steps · CIFAR100-B50-5steps | Avg accuracy (%): 73.21 -> 74.50 | Helped |
Surprising findings
- Removing knowledge distillation slightly increases Avg (75.51 vs 75.42), indicating KD gives no significant gain when combined with other components.
- Applying adaptive logit alignment during feature-extractor training gave insufficient benefit, so it is only applied in the classifier fine-tuning stage.
Limitations
- Last-step accuracy lags ANN methods, especially at 20 steps (59.43% vs DNE 67.61%), and the gap widens with more steps.
- Small time window (T=4) is said to limit SNN fitting and generalization ability.
- Energy efficiency is only a theoretical estimate, not measured on hardware.
Experimental setup
Datasets: CIFAR100 DVS-CIFAR10
Benchmarks: CIFAR100-B0 5/10/20 steps CIFAR100-B50 5/10 steps DVS-CIFAR10
Metrics: average incremental accuracy (Avg) last-step accuracy (Last)
Baselines: DER Simple-DER iCaRL UCIR BiC WA PODNet RPSNet Dytox+ TCIL DNE DSD-SNN
Builds on: DER TET loss iCaRL
Closest papers by meaning
- Temporal Flexibility in Spiking Neural Networks: Towards Generalization Across Time Steps and Deployment Friendliness (ICLR 2025 · similarity 0.87)
- Efficient Logit-based Knowledge Distillation of Deep Spiking Neural Networks for Full-Range Timestep Deployment (ICML 2025 · similarity 0.87)
- Efficient Spiking Neural Networks with Sparse Selective Activation for Continual Learning (AAAI 2024 · similarity 0.87)
- Adaptive Calibration: A Unified Conversion Framework of Spiking Neural Networks (AAAI 2025 · similarity 0.86)
- Temporal Separation with Entropy Regularization for Knowledge Distillation in Spiking Neural Networks (CVPR 2025 · similarity 0.86)
- Synergy Between the Strong and the Weak: Spiking Neural Networks are Inherently Self-Distillers (NeurIPS 2025 · similarity 0.86)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.