Research map / Probabilistic and scientific modeling / Dynamical systems
MA-Net: Rethinking Neural Unit in the Light of Astrocytes
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 MA-Net, a neural network whose connections are adaptively modulated during training by a multi-astrocyte-neuron (MA-N) model with temporal and joint mechanisms, reducing parameters while improving classification, segmentation and detection.
Problem
Standard artificial neuron networks have fixed structures during training, leading to overfitting and over-parameterization; existing astrocyte-inspired models use only a single astrocyte with one state.
Approach
Models multiple sub-astrocytes that modulate neuron-to-neuron weights bidirectionally, with temporal (LSTM-based, history-dependent) and joint (chain-optimized across astrocytes) modulation, and an alternating optimization between a task network and a modulation network that prunes weak connections.
Key results
- On CIFAR10, MA-Net (RNT50) reaches 97.10% accuracy with 7.36M params vs. 96.78% for CDS at 25.60M params.
- On ImageNet-1k, Ours (RNT50) reaches 78.96% top-1 with 8.16M params vs. 78.31% for AstroNet at 8.70M params.
- On COCO with CondInst, the RNT50 backbone pre-trained with MA-Net achieves 40.7 AP vs. 40.0 for AstroNet.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Replace the RNT18 backbone with MA-Net connection optimization | RNT18 · CIFAR10 | accuracy: 96.52 -> 96.81 | Helped |
| Replace the RNT50 backbone with MA-Net connection optimization | RNT50 · CIFAR10 | accuracy: 96.76 -> 97.10 | Helped |
| Replace the RNT18 backbone with MA-Net connection optimization | RNT18 · ImageNet-1k | accuracy: 72.85 -> 73.27 | Helped |
| Replace the RNT50 backbone with MA-Net connection optimization | RNT50 · ImageNet-1k | accuracy: 78.31 -> 78.96 | Helped |
| Use MA-Net-pretrained backbone instead of AstroNet-pretrained backbone | BoxInst on COCO · COCO segmentation (BoxInst) | AP: 32.7 -> 34.7 | Helped |
| Use MA-Net-pretrained backbone instead of AstroNet-pretrained backbone | CondInst on COCO · COCO segmentation (CondInst) | AP: 40.0 -> 40.7 | Helped |
| Use MA-Net-pretrained backbone instead of AstroNet-pretrained backbone | RCNN on COCO · COCO object detection (RCNN) | AP: 38.0 -> 39.8 | Helped |
| Use MA-Net-pretrained backbone instead of CDS-pretrained backbone | DETR on COCO · COCO object detection (DETR) | AP: 43.2 -> 44.0 | Helped |
| Apply MA-Net connection optimization to ViT | ViT · ImageNet-1k | accuracy: 77.90 -> 79.47 | Helped |
| Apply MA-Net connection optimization to Swin | Swin · ImageNet-1k | accuracy: 84.50 -> 85.66 | Helped |
| Use MA-Net instead of Boost-DyNN | RNT50 · ImageNet-1k | accuracy: 76.05 -> 78.96 | Helped |
| Use 4 astrocytes instead of 2 (ablation) | MA-Net with RNT18 · CIFAR10 | accuracy: 96.70 -> 96.81 | Helped |
Limitations
- Paper states reported figures are averages of five repeated runs, but no variance is reported
- Pruning threshold and astrocyte count trade off accuracy against time cost
- Gains over strongest baselines are small (often under 1%)
Experimental setup
Datasets: CIFAR10 ImageNet-1k COCO
Benchmarks: CIFAR10 ImageNet-1k COCO
Metrics: accuracy AP parameters (M) time cost (GPU days)
Baselines: AstroNet CDS NAS-LID HOTNAS Boost-DyNN BoxInst CondInst ViT Swin
Builds on: AstroNet CDS Vision Transformer Squeeze-and-Excitation
Closest papers by meaning
- GliaNet: Adaptive Neural Network Structure Learning with Glia-Driven (CVPR 2025 · similarity 0.88)
- The Expressive Leaky Memory Neuron: an Efficient and Expressive Phenomenological Neuron Model Can Solve Long-Horizon Tasks. (ICLR 2024 · similarity 0.83)
- Scaling Laws and Tradeoffs in Recurrent Networks of Expressive Neurons (NeurIPS 2026 · similarity 0.83)
- Feedback control guides credit assignment in recurrent neural networks (NeurIPS 2024 · similarity 0.83)
- Shaping the distribution of neural responses with interneurons in a recurrent circuit model (NeurIPS 2024 · similarity 0.83)
- Structured flexibility in recurrent neural networks via neuromodulation (NeurIPS 2024 · similarity 0.82)
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.