IC-293ViT-B/16 and ViT-B/32 are largely redundant: retaining only top-5 neurons per layer while zeroing all others preserves most classification accuracy

Yifan Wang, Yifei Liu, Yingdong Shi, Changming Li, Anqi Pang, Sibei Yang, Jingyi Yu, Kan Ren

SourceDiscovering Influential Neuron Path in Vision Transformers

The paper prunes neurons by zeroing out all FFN neurons except the top-t most frequently selected by the neuron path method (t in {1,5,10,30,50}), tested on held-out 20% of the validation set. For ViT-B/16, retaining top-5 neurons per layer while pruning 100% of other neurons maintains accuracy near the ~69% baseline. For ViT-B/32, the same procedure keeps accuracy near the ~53% baseline. Notably, increasing the number of retained neurons beyond 5 leads to a decline in performance, suggesting that non-path neurons can be harmful. A comparison with ViT-Slim (restricted to FFNs) shows JAS-prune achieving 67.7% vs ViT-Slim's 9.96%.

Evidence
interventional
Key metric
ViT-Slim accuracy 9.96% vs JAS-prune 67.7% (Table 2). Ablation (Table 3, 100% mask): JAS topk=5 accuracy 0.521252 (depth 6), 0.527151 (depth 9), 0.540702 (depth 12) vs activation topk=5: 0.517991, 0.518137, 0.531976.
Caveat
The pruning is class-specific (top neurons identified per class from 80% of data, applied to 20% held-out). The paper notes this is a generalization setting but the class-specificity limits direct applicability to a single global pruning.
Model
ViT ViT-B/16, ViT-B/32
Datasets
ImageNet-1k / ImageNet / ImageNet-1k-val / ImageNet-Val [eval]
Methods
ViT-Slim [compared-to]
Related work
Mixture-of-Depths [context]
Related findings
IC-290, IC-291, IC-292
Extraction
automatic-extraction