If I have 5 separate binary classifiers that are using a pre-trained Inception V3 each separately would it provide less accurate results if I modify an inception V3 to create multi-output results as shown in the paper below?
Is there a scientific proof or article that shows multioutput classification of binary classification tasks produces better results than separately considering each as a binary task? Is this because of weight sharing?
Here’s the article I saw this in:
Predicting gene mutational status from whole-slide images. We next focused on the LUAD slides and tested whether CNNs can be trained to predict gene mutations using images as the only input. For this purpose, gene mutation data for matched patient samples were downloaded from TCGA. To make sure the training and test sets contained enough images from the mutated genes, we only selected those which were mutated in at least 10% of the available tumors. From each LUAD slide, only tiles classified as LUAD by our classification model were used for this task in order to avoid biasing the network to learn LUAD-specific versus LUSC-specific mutations and to focus instead on distinguishing mutations relying exclusively on LUAD tiles. Inception v3 was modified to allow multioutput classification (Methods): training and validation was conducted on ~212,000 tiles from ~320 slides, and testing was performed on ~44,000 tiles from 62 slides. Box plot and ROC curve analysis (Fig. 3a,b and Supplementary Fig. 5) show that six
frequently mutated genes seem predictable using our deep-learning approach; AUC values for serine/threonine protein kinase 11 (STK11), EGFR, FAT atypical cadherin 1 (FAT1), SET binding protein 1 (SETBP1), KRAS and TP53 were between 0.733 and 0.856 (Table 1). Availability of more data for training is expected to substantially improve the performance.