Are the High-weight Neurons the Important Ones in Image Classification Neural Networks?

This paper studies whether large-weight neurons are truly important in image classification networks, using experiments on CIFAR-10 and Mini-ImageNet. The top 10% high-weight neurons overlap with accuracy-impacting neurons by at most about 25%. Perturbing them can reduce accuracy by 45–80% under certain operations, compared with 3–7% for random perturbations, yet roughly one third have minimal impact. Removing the top 10% and retraining leaves accuracy 10–20% below baseline, while removing only the top 0.1% permits near-full recovery. Some low-weight intervals also cause 10–17% degradation when perturbed.