Research map / Label-efficient and robust learning / Zero-/few-shot and meta-learning

Manhattan Self-Attention Diffusion Residual Networks with Dynamic Bias Rectification for BCI-based Few-Shot Learning

AAAI 2025 · Label-efficient and robust learning › Zero-/few-shot and meta-learning · approach: Fine-grained and cross-domain few-shot baselines · 1 citations (OpenAlex)

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 MSADiff-Resnet, a ResNet with Manhattan self-attention diffusion layers plus dynamic inter/intra-class bias rectification, for few-shot EEG/BCI classification on a multi-source dataset built from BCICIV-2a, SEED-V and BETA.

Problem

Few-shot learning on BCI/EEG data suffers from distribution biases across multi-source datasets and scarce labeled support samples, and prior diffusion-based FSL uses fixed diffusion intensity and ignores biases and unlabeled data.

Approach

Adds a Manhattan self-attention diffusion layer (attention weights times a Manhattan-distance decay) to ResNet blocks to control local diffusion and a global diffusion strength, combined with inter-class shift correction and attention-weighted intra-class prototype rectification with pseudo-labeled query samples.

Key results

• On the multi-source BCI dataset with ResNet18, the proposed method reaches 55.49% (5-way 1-shot) and 75.48% (5-way 5-shot) on BCICIV-2a, outperforming the best baseline by about 2 points.
• Ablation on the multi-source dataset (Conv4, 1-shot): full model 37.71% vs 36.06% with MSADiff only and 35.93% with dynamic bias rectification only.
• Diffusion layer count r=6 selected as optimal; accuracy plateaus beyond 6 layers with gamma=0.5.

Measured effects reported in the paper

Limitations

• Excessively high global diffusion intensity leads to anomalous results, requiring careful tuning of gamma.
• Evaluated only on EEG-derived GAF images from three datasets; no temporal information integrated.
• Improvements over strongest baselines are modest (about 1-3%) and the bias rectification analysis is largely empirical.

Experimental setup

Datasets: BCICIV-2a SEED-V BETA

Benchmarks: BCICIV-2a SEED-V BETA multi-source dataset

Metrics: accuracy standard deviation p-value

Baselines: PN MSCNN BD-CSPN Meta-DM AFR DA RelationNet Diff-ResNet DiffAlign

Builds on: Diff-ResNet BD-CSPN Manhattan self-attention Prototypical Networks DiffAlign

Closest papers by meaning

• An Efficient Framework for Enhancing Discriminative Models via Diffusion Techniques (AAAI 2025 · similarity 0.84)
• Neural Residual Diffusion Models for Deep Scalable Vision Generation (NeurIPS 2024 · similarity 0.83)
• Self-Attention-Based Contextual Modulation Improves Neural System Identification (ICLR 2025 · similarity 0.83)
• Denoising Task Routing for Diffusion Models (ICLR 2024 · similarity 0.83)
• Diff-Plugin: Revitalizing Details for Diffusion-based Low-level Tasks (CVPR 2024 · similarity 0.83)
• Faster Video Diffusion with Trainable Sparse Attention (NeurIPS 2025 · similarity 0.83)