Vanishing Feature: Diagnosing Model Merging and Beyond

Xingyu Qu, Samuel Horváth

TL;DR

Merging neural networks trained from different initializations often leads to catastrophic performance drops. This paper identifies the root culprit as the "vanishing feature" phenomenon—where input-dependent representations decay across layers while input-independent bias terms dominate—and introduces Preserve-First Merging (PFM) alongside refined normalization strategies to match or beat edge models without retraining. The authors also show this same phenomenon undermines one-shot model pruning and can be largely salvaged using the same post-hoc normalization lens.

What Problem Does This Paper Address?

Model merging aims to combine the weights of independently trained neural networks into a single model, reaping the ensemble-like benefits of multiple models without the extra inference or storage overhead. While merging fine-tuned checkpoints from the same pre-trained model works well, merging models trained from different random initializations routinely collapses to random-guess accuracy.

Prior work attributed this failure to permutation invariance (neurons in different models are ordered differently) and "variance collapse" (activation variance shrinks in the merged model). However, existing solutions like permutation matching and activation normalization (e.g., REPAIR) often require re-training, huge model widths, or yield inconsistent results on deeper architectures. The fundamental mechanics of why features break down during propagation remained unclear.

Methods

  • Feature-Bias Decomposition: For a network of depth $L$, intermediate pre-activations are decomposed into an input-induced latent feature $g^{(l)}(x)$ and an input-independent offset $h^{(l)}$ driven by parameter biases: $$f^{(l)}(x) = g^{(l)}(x) + h^{(l)}$$
  • Enhanced Normalization (RESCALE & Bias Calibration): Standard REPAIR normalizes activations and re-injects the edge models' expected offset $\mathbb{E}[f(x)]$. When $g^{(l)}(x)$ has vanished, re-injecting offsets can cement bias dominance. The authors propose:
  • RESCALE: Normalizes by merged variance and rescales by target variance without shifting the mean.
  • Bias Calibration / Backward Removal: Scaling or zeroing out biases in deeper layers so they cannot overwhelm the dying input signal.
  • Preserve-First Merging (PFM): Finding the optimal global permutation is NP-hard, and early feature mismatch compounds through the network. PFM keeps the first $l$ layers unmerged (running inputs through separate early backbones) before fusing intermediate representations with permutation matching and merging all remaining layers.
  • Post-Pruning Normalization: One-shot pruned models also suffer from vanishing features. Applying REPAIR or RESET using statistics from the dense parent model restores activation variance without retraining or extra parameters.
       Preserve-First Merging (PFM)
Input ──┬──> [Model 0: Layers 1..l] ──┐
        └──> [Model 1: Layers 1..l] ──┴──> [Permute & Merge] ──> [Merged Layers (l+1)..L] ──> Output

Datasets

  • Vision Benchmarks: MNIST, CIFAR-10, CIFAR-100, and ImageNet (ILSVRC 2012).
  • Architectures: Multi-layer linear networks, VGG16 (standard, BatchNorm, LayerNorm variants), ResNet20, ResNet50, ConvNeXt (T and B variants), and Vision Transformers (ViT).
  • Note: No new datasets are introduced; the paper evaluates standard computer vision benchmarks.

Key Insights and Ideas

  • Biases Drown Features in Deeper Layers: Merging weights with conflicting signs shrinks parameter magnitudes. In linear networks, weight shrinkage compounds multiplicatively across $l$ layers ($O(\lambda^l)$ for input features), whereas biases introduced at layer $j$ decay only over $l-j$ layers. Consequently, deep layers become completely detached from input data $x$ and output constant noise.
  • Residual Connections Rescue Merging: Skip connections preserve lower-order matrix products across layers, which inherently mitigates feature vanishing. This provides a theoretical explanation for why ResNets are significantly easier to merge than plain CNNs.
  • Training Hyperparameters Dictate Mergability: Training edge models with a higher learning rate (e.g., 0.1 vs. 0.01) and stronger weight decay introduces implicit regularization that contracts the parameter distance between distinct optima, reducing post-merging loss barriers by over 40–60%.
  • Early Layers Are Non-Negotiable: Errors in early layers compound irreversibly. Preserving even a fraction of the early layers unmerged prevents catastrophic feature collapse.

Key Results

  • CIFAR-10 VGG16 Surpasses Edge Models: By preserving the first 8 of 14 layers ($\text{PFM}_{8/14}$) and applying activation matching with RESCALE, a merged VGG16 achieved 92.43% test accuracy on CIFAR-10—beating both individual edge models (~91.7% and 91.8%) without any post-training, with only a 20% parameter increase.
  • RESCALE vs. REPAIR: On VGG16 with ZipIt! merging, RESCALE improved midpoint accuracy to 84.70% compared to REPAIR’s 67.93%.
  • Pruning Recovery: For an 80% sparse ConvNeXt-B on ImageNet pruned via simple magnitude pruning, applying post-pruning REPAIR boosted top-1 accuracy from 35.45% to 69.88%. Remarkably, estimating BatchNorm statistics with as few as 64 calibration samples was sufficient.

Limitations and Future Work

  • Compute/Parameter Trade-Off: PFM keeps early layers duplicate-branched, increasing parameter count and FLOPs proportionally to the number of preserved layers (e.g., 1.2–1.8× compute/parameter overhead).
  • Scaling to Complex Datasets: On ImageNet ResNet50, PFM requires preserving a much larger portion of layers to match edge performance, showing that existing normalization techniques still struggle with deep, highly complex representations.
  • Theoretical Scope: Theoretical proofs for feature vanishing are derived primarily on linear networks; extending rigorous proofs to non-linear activations with normalization layers remains open for future research.

Summary updated

Abstract

Model merging offers an efficient way to combine pre-trained neural networks but often suffers from inconsistent performance, especially when merging models with different initializations. We identify the ''vanishing feature'' phenomenon, where input-induced features diminish during propagation through the merged model, degrading performance. Through theoretical and empirical analysis, we reveal that this phenomenon underpins challenges like variance collapse and explains techniques like permutation-based merging, post-merging normalization, etc. We show that existing normalization strategies can be enhanced by precisely targeting the vanishing feature issue. Leveraging these insights, we propose the ''Preserve-First Merging'' (PFM) strategy, which preserves early-layer features, enabling merged VGG16 models on CIFAR-10 to surpass the original models without post-training for the first time. Furthermore, we demonstrate that the vanishing feature phenomenon extends to other contexts, such as model pruning. Applying post-pruning normalization to mitigate the issue significantly improves one-shot pruning performance at high sparsity, offering a simple and effective post-pruning solution. The code is available at https://github.com/XingyuQu/VF.
BibTeX
@inproceedings{qu2025vanishing,
  author = {Xingyu Qu and Samuel Horváth},
  title = {{Vanishing Feature: Diagnosing Model Merging and Beyond}},
  booktitle = {CPAL (Proceedings Track)},
  year = {2025},
  url = {https://openreview.net/forum?id=KBSh5ChQIo},
}
Figure