Reproduce “Demystifying When Pruning Works via Representation Hierarchies” on OpenResearch

About this paper

Shwai He, Guoheng Sun, Haichao Zhang, Yun Fu, Ang Li

This research explains the inconsistent effectiveness of network pruning across generative and non-generative language tasks by analyzing its impact on internal representation hierarchies. It demonstrates that while embedding and logit spaces remain robust, the nonlinear softmax function amplifies pruning-induced perturbations in the probability space, leading to severe performance degradation in generative tasks due to error accumulation, while non-generative tasks are more resilient.

Key insights

• Embedding and logit spaces are highly robust to pruning-induced perturbations, maintaining high cosine similarity with original representations, suggesting core semantic information is largely preserved.
• The nonlinear softmax function significantly amplifies subtle logit perturbations into substantial deviations in the probability space, making generative tasks highly sensitive to pruning.
• Generative tasks are particularly vulnerable due to the accumulation and propagation of these amplified errors during autoregressive decoding, whereas non-generative tasks benefit from the stability of relevant probability subspaces where decision-making remains robust.

Summary by alphaXiv.