I wish to make an ensemble of deep models to solve a classification problem. I want to know how many models shall be used to create that ensemble to ensure unbiased results. I have head 30+ models shall be used to determine the skill of the model but cannot find anything relating to ensemble modelling.
There are no restrictions/guidelines on the number of models. You can start even from 3 models. You can keep the number of models as a hyperparameter if the training cost is less. Typically, you will observe the slanted 'L' shaped curve for MSE vs # of models plot. You can take the elbow point as the final # of models.