Research map / Architectures and efficiency / CNN design
Model LEGO: Creating Models Like Disassembling and Assembling Building Blocks
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 Model Disassembling and Assembling (MDA), which extracts task-aware components from trained CNN classifiers via relative-contribution-based locating and reassembles them into new task-specific models without retraining.
Problem
Training new deep models is resource-intensive; the paper asks whether task-specific models can be created by disassembling and recombining parts of already-trained networks without training.
Approach
Define relative contribution of feature maps and filters through contribution aggregation and allocation across layers, locate task-relevant filters/kernels/biases via thresholds, then assemble components from different isomorphic CNNs using alignment padding (empty kernels) and parameter scaling of filters.
Key results
- Disassembled single-category components reach 100% accuracy on CIFAR-10 and CIFAR-100 single tasks, vs. base accuracies of 84-96%, with multi-task disassembly often within a few points of or above the source model.
- Assembled models without fine-tuning match or exceed source accuracy in several settings (e.g., GoogleNet on CIFAR-10+Tiny-ImageNet task 0-2+0-69: 57.51% vs base 54.59%), but drop notably for larger combinations (e.g., CIFAR-100+Tiny-ImageNet 0+0 on VGG-16: 50.00% without fine-tuning).
- Parameter scaling improves the '0 + 0' CIFAR-10+CIFAR-100 VGG-16 assembly from 50.00% (padding only) to 87.00%.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Disassemble task-specific components with MDA rather than use the source model as-is | CNN classifiers · CIFAR-100 task 0-19, GoogleNet | accuracy: 75.90 -> 87.55 | Helped |
| Disassemble task-specific components with MDA rather than use the source model as-is | CNN classifiers · CIFAR-10 task 3-9, ResNet-50 | accuracy: 93.81 -> 94.46 | Helped |
| Disassemble task-specific components with MDA rather than use the source model as-is | GCN · Cora task 0 | accuracy: 72.36 -> 100.00 | Helped |
| Assemble task-aware components from separate datasets into one model | GoogleNet · CIFAR-10 + Tiny-ImageNet, tasks 0-2 + 0-69 | accuracy: 54.59 -> 57.51 | Helped |
| Assemble task-aware components from separate datasets into one model | ResNet-50 · CIFAR-100 + Tiny-ImageNet, tasks 20-69 + 70-179 | accuracy: 76.54 -> 56.53 | Hurt |
| Add parameter scaling to padding alignment (ablation) | VGG-16 assembly · CIFAR-10 + CIFAR-100, tasks 0 + 0 | accuracy: 50.00 -> 87.00; 37.00% | Helped |
| Add parameter scaling to padding alignment (ablation) | VGG-16 assembly · CIFAR-10 + CIFAR-100, tasks 0-2 + 0-19 | accuracy: 71.74 -> 74.17 | Helped |
| Add parameter scaling to padding alignment (ablation) | VGG-16 assembly · CIFAR-10 + CIFAR-100, tasks 3-9 + 20-69 | accuracy: 71.97 -> 73.72 | Helped |
| Add parameter scaling to padding alignment (ablation) | VGG-16 assembly · CIFAR-10 + CIFAR-100, tasks 0-9 + 20-99 | accuracy: 69.12 -> 72.07 | Helped |
| Increase the component-locating thresholds α and β (ablation) | MDA disassembly · CNN classifier datasets in threshold ablation | accuracy | Hurt |
| Disassemble more layers, from deep layers toward shallow layers (ablation) | MDA disassembly · CNN classifier datasets in layer-count ablation | accuracy | Mixed |
Surprising findings
- Disassembled components often surpass the accuracy of the source model on their assigned tasks, despite no retraining.
- Assembled models with more combined categories can fall below the source model, which the authors attribute to interference between components.
Limitations
- Assembled models can lose accuracy when many task components are combined, due to interference from irrelevant components.
- Restricted to CNN classifiers and isomorphic architectures; GCN and other architectures only briefly tested.
- Compression results have higher FLOPs and parameters than pruning baselines such as FPGM and HRank.
Experimental setup
Datasets: CIFAR-10 CIFAR-100 Tiny-ImageNet MNIST FASHION-MNIST ImageNet Cora
Benchmarks: CIFAR-10 CIFAR-100 Tiny-ImageNet ImageNet Cora MNIST FASHION-MNIST
Metrics: accuracy FLOPs parameter count
Baselines: FPGM HRank KA (Knowledge Amalgamation) GCN
Builds on: Neural Architecture Disentanglement Layer-wise Relevance Propagation Knowledge Amalgamation
Closest papers by meaning
- ZipIt! Merging Models from Different Tasks without Training (ICLR 2024 · similarity 0.86)
- Merging on the Fly Without Retraining: A Sequential Approach to Scalable Continual Model Merging (NeurIPS 2025 · similarity 0.85)
- Neural Assembler: Learning to Generate Fine-Grained Robotic Assembly Instructions from Multi-View Images (AAAI 2025 · similarity 0.84)
- DD-CAM: Minimal Sufficient Explanations for Vision Models Using Delta Debugging (NeurIPS 2026 · similarity 0.84)
- Dataset Distillation for Pre-Trained Self-Supervised Vision Models (NeurIPS 2025 · similarity 0.84)
- REPA Works Until It Doesn’t: Early-Stopped, Holistic Alignment Supercharges Diffusion Training (NeurIPS 2025 · similarity 0.84)
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.