4 ms·
The primary thing missing from all of these guides is that you need to have two things for ML: 1. A purpose for utilizing it 2. A data set to train/act on Wi
by AndrewKemendo 10y ago
The primary thing missing from all of these guides is that you need to have two things for ML:
1. A purpose for utilizing it
2. A data set to train/act on
Without that, all you get are a bunch of shovels and picks, but no idea of what kind of wood/bricks you need or a plan for the house.
- estsauver 10y agoI 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.
- hrant25 10y agoThanks for the feedback, Andrew! This guide is designed mostly for those who already know what problems they want to solve, but don't know how to start or where to look for high quality and up-to-date educational resources. Also this is designed for those who want to do research in this area and just want to develop better shovels (that was me 2 years ago). Regarding datasets, each course we suggest in the guide has its own way of dealing with datasets. Most of them teach how to work with MNIST, which is pretty good for many purposes. I agree that another guide on datasets could be useful for some people.
- AndrewKemendo 10y agoYes, thank you for the resources and my comment is not an indictment on your page as I think it's well organized. Moreso I'm commenting broadly on the recurrence of similar guides and how they are generally not accessible to people with no exposure to ML. All the best.