Multiple Context Learning Networks for Visual Question Answering

Table 3

The results of fine-tuning BERT on VQA v2.0 validation set.

Model	lr×	All	Y/N	Num	Other
N = 1, BERT	0.001	65.21	82.61	45.74	57.13
N = 1, BERT	0.01	66.12	84.08	45.97	57.80
N = 1, BERT	0.1	66.61	85.09	46.14	57.98
N = 2, BERT	0.1	67.52	85.18	49.09	58.97
N = 3, BERT	0.1	67.86	85.35	49.75	59.27
N = 4, BERT	0.1	67.80	85.73	49.94	58.94
N = 3, LSTM	-	66.86	84.62	48.85	58.34