playlist

Well, this 10. But in [snorts] [snorts] It's 2:00. Let's just keep on rocking . I think that microphone gets weaker and weaker . Anyway, as long as it's on the recording, it's all fine . So ladies and gents, um last time we left off with this repetition of classic machine learning and um we investigated linear regression, right ? So the usual model uh is that right ? So this is a linear model meaning you take a weight vector you take a vector of observations and this is your model right that you use right linear regression means uh you model it like this right so if I've absorbed the bias already right this is easy to do by saying that the uh zeroth weight is is one right or the zero uh feature uh value is one um which has some performance benefits compared to um the other sort of classifier we'll look at in a second . Right? So um logistic uh regression or a one layer neural network, right ? at least conventional feed forward neural network um which you can stack and you get deep learning . Okay. Um . Um if you need to of course you are aware that we can substitute x with five of x . So that's a feature map right ? So this is the raw features and this is the mapped features right . So you can in that regard you would have uh linear regression as that and we know for a one-dimensional x we can take powers of x as we see fit right so this is that what you can also do with these linear models is kernelize it this is also well known from machine learning one so the solution for this there's a closed form solution um more penro sudtor inverse of the um feature matrix , right? Um and then you you'd have your weights . Okay? H in this equation what you see is you get um inner products of that or its feature version which you can replace by a kernel function