3 ms·
It's probably more training-compute intensive, but they can do drop-out, right? The strategy they used for ImageNet recognition, when they were using supervised
by anabis 3y ago
It's probably more training-compute intensive, but they can do drop-out, right?
The strategy they used for ImageNet recognition, when they were using supervised learning and training data was scarse.
- minimaxir 3y agoDropout is one strategy for regularization but doesn't guarantee avoiding overfitting, especially now that modern AI models generalize much better than they did during the ImageNet days. Many of the big LLMs use a dropout of 0.1 though.