Fig. 19
From: Towards a robust out-of-the-box neural network model for genomic data
Precision-recall curves of LSTM-layer model with two different optimizers: Adam and SGD on three datasets of increasing size. The higher the curve, the better performance with a horizontal dashed line to represent random prediction. An ideal precision-recall curve would cross the (1,1) point. The splice data has three panels since precision-recall curves assume binary classification and the splice dataset has three classes (0, 1, 2). Each panel corresponds to prediction one class vs the other two combined. The best optimizer varies with SGD outperforming Adam for the smallest dataset (splice) and Adam outperforming SGD on the other two datasets. It is widely accepted that SGD performs better in terms of finding global optima. However, due to its low speed, it can get stuck in one plateau too long