The history of ImageNet

Good read.

> realized that data was such an important piece of the development of Machine Learning algorithms early on, whereas most of her colleagues did not believe the same.

I think that it was hard to deny, even before ImageNet, that more data gives better models. But maybe people didn't believe that the return on investment in data would be that high?

We are currently reaping the low hanging fruit of CNNs and annotated data collection. But this approach has diminishing returns: you need to collect an order of magnitude more annotations to get slightly improved accuracy, and the models become exponentially larger.

More annotations enabled an annotations hungry method (Deep Learning) to dominate. But we know that brains do not train on an exponential amount of annotations, but on a very limited amount of data.

Shouldn't residual blocks get a mention here?