2 ms·
I will mention that you can get surprisingly far with just a little bit of labeled training data now. Transfer learning is a pretty powerful technique.
by estsauver 10y ago
I will mention that you can get surprisingly far with just a little bit of labeled training data now. Transfer learning is a pretty powerful technique.
- 1024core 10y agoYou still need to train a model on what you're transferring from. It's not like you can manufacture trained models out of thin air.
- AlexCoventry 10y agoI think the idea is that, for example, you start with some pre-trained visual object-identification network, remove the final logistic-regression layer, and replace with a layer which you train your problem on.
- AndrewKemendo 10y agoUndoubtedly. However there are three sets of data you need to a functional system: Training Data and Validation data sets (a lot of sources exist already) and the data set you want to get insights from (Novel data sets usually). These of course work together in reinforcement scenarios. So for example if you want to build a tool that tells you what kind of objects are in any given image you can use ImageNet to Train and Validate your CNN. At that point you have a classifier. But for an actual useful application you need a user to submit new images to actually give some kind of valuable output. It's that second data set that is super hard to get - because it's novel.