 So now we finally get to domain adaptation, which is actually something that suddenly in research But also for most application you will be doing a lot. What is the typical setting? We have a data set that we're interested in that data set is Big, but it's not huge So say if we work in the medical domain, we might be able to take 1,000 photographs of some medical pram we will not be able to take 10 million photographs of that medical pram and That's so we have our own data set, which is not huge But there's another data set that is really really big somewhere out there on the internet which is somewhat similar and And somewhat similar could be I have a photo of a potential tumor and I still use image net because image net I images that are made in like a slightly similar way to tumor enough similar enough that it's useful to use that so Then what is the idea? I first take a network and train it on a really big data set in a way we're setting a prior if you thought a lot a bit Bayesian ideas and Only after first training on the big data set we then train of our data set of interest Now it's even simpler in practice because for most big data sets We can actually download the already trained network. So this first step has already been done for us We just now need to retrain it on the data that we have and of course as with most domains There's lots of ways of doing this a little bit better But it is this very idea that I take one data set for which I have a lot I train on that and I use that to start my training on the data set that I'm actually interested in and that usually gives massive advantages and Why do you want to do that? Well, usually you simply don't have enough money to even train a big network Like it's gotten the the really big image recognition and video data sets and text data sets They're all so big that the training would set you back rather serious Amounts of money whereas just retraining it to solve your specific prom will be relatively efficient and In practice, you can also often not allow yourself to label enough data No, it's expensive and if you want a really deep really rich neural network of the Of the object recognition variety, for example You you will simply not be able to label the millions of images that you'd otherwise need So in practice, you almost always need to do domain adaptation and it buys you a lot of mileage So let's say you want to classify Pokemon. How should we start? Well, why don't you think a little about that during the domain adaptation exercise?