3 ms·
Lately there have been amazing results on pairs of web text+image, no need for labelling. These datasets are hundreds of times larger and cover many more catego
by visarga 5y ago
Lately there have been amazing results on pairs of web text+image, no need for labelling. These datasets are hundreds of times larger and cover many more categories of objects. GPT-3 is also trained on raw text. I think ImageNet and its generation have become toy datasets by now.