Table 4 Ablation experimental results.

Model	All F.		Selected F.		Easy F.		Biological F.		Features				100%	10%	1%
Model	\(100\%\)	\(10\%\)	\(100\%\)	\(10\%\)	\(100\%\)	\(10\%\)	\(100\%\)	\(10\%\)	All	Selected	Easy	Biological	100%	10%	1%
IPDM	87.80	87.30	87.55	87.20	80.05	77.45	68.60	65.35	88.75	85.00	85.00	80.00	80.70	80.61	79.48
w/o Input prompt	44.75	44.75	44.75	44.75	44.75	44.75	44.75	44.75	65.00	65.00	65.00	65.00	75.43	75.43	75.43
w/o Prompt-tuning	87.65	86.35	87.50	87.15	79.15	74.30	67.60	64.55	78.75	82.50	82.50	80.00	80.47	80.66	78.40
w/o MoE	87.75	87.10	87.50	86.60	80.00	76.90	67.60	65.15	85.00	85.00	85.00	78.75	80.73	80.66	79.06
w/o Prompt-tuning &MoE	87.45	76.05	87.50	86.40	78.45	73.40	66.20	63.75	77.50	77.50	82.50	78.75	80.22	80.39	78.39

The structure of table is consistent with Table 3, “F.” is the abbreviation of “Features”. “w/o Input Prompt” means no meta-information is used to construct input prompts, “w/o Prompt-tuning” means traditional fine-tuning, “w/o MoE” means only one expert, and “w/o Promt-tuning & MoE” means traditional fine-tuning and one expert both. The results of “w/o Input Prompt” overfits the class of the highest proportion. Either “w/o Prompt-tuning” or “w/o MoE” only has a small drop in accuracy for most settings, but it becomes obvious when “w/o Promt-tuning & MoE”.

Quick links

Search