Pruning small-magnitude weights from pre-trained LLMs causes monotonic irreversible performance degradation on difficult downstream tasks, supporting the Junk DNA Hypothesis that these weights hold essential knowledge.
Deep compression: Compressing deep neural networks with pruning, trained quantization and huffman coding
1 Pith paper cite this work. Polarity classification is still indexing.
1
Pith paper citing it
fields
cs.LG 1years
2023 1verdicts
UNVERDICTED 1representative citing papers
citing papers explorer
-
Junk DNA Hypothesis: Pruning Small Pre-Trained Weights Irreversibly and Monotonically Impairs "Difficult" Downstream Tasks in LLMs
Pruning small-magnitude weights from pre-trained LLMs causes monotonic irreversible performance degradation on difficult downstream tasks, supporting the Junk DNA Hypothesis that these weights hold essential knowledge.