- Видео 46
- Просмотров 17 601
Richard Turner
Добавлен 17 апр 2010
Видео
1.6 Types of inference problem and an outline of the course
Просмотров 2593 года назад
1.6 Types of inference problem and an outline of the course
1.3.1 A little more information about Cox's theorem
Просмотров 5243 года назад
For more background about the probabilistic approach to machine learning and statistics, please see this blog article: mlg-blog.com/2021/03/31/what-keeps-a-bayesian-awake-at-night-part-1.html
5.4 The K-Means Algorithm: Limitations
Просмотров 2413 года назад
5.4 The K-Means Algorithm: Limitations
1.2 The radioactive decay problem and ad hoc solutions
Просмотров 2923 года назад
1.2 The radioactive decay problem and ad hoc solutions
1.4 The radioactive decay problem: improving over ad hoc solutions and visualising the solution
Просмотров 2343 года назад
1.4 The radioactive decay problem: improving over ad hoc solutions and visualising the solution
5.6 Mixtures of Gaussians: Parameter Learning
Просмотров 2433 года назад
5.6 Mixtures of Gaussians: Parameter Learning
5.5 The Probabilistic Approach to Clustering
Просмотров 1,4 тыс.3 года назад
5.5 The Probabilistic Approach to Clustering
1.3 The radioactive decay problem: a principled approach
Просмотров 2293 года назад
1.3 The radioactive decay problem: a principled approach
6.15 Maximum likelihood learning of HMMs
Просмотров 4633 года назад
6.15 Maximum likelihood learning of HMMs
6.10 Detailed derivation of general filtering equations
Просмотров 1863 года назад
6.10 Detailed derivation of general filtering equations
6.9 Overview of the general Kalman filtering recursion
Просмотров 2993 года назад
6.9 Overview of the general Kalman filtering recursion
6.14 Computing the likelihood of the parameters in HMMs
Просмотров 1463 года назад
6.14 Computing the likelihood of the parameters in HMMs
6.1 Introduction to Sequence Modelling
Просмотров 1613 года назад
6.1 Introduction to Sequence Modelling
5.9 The Expectation Maximisation Algorithm
Просмотров 3133 года назад
5.9 The Expectation Maximisation Algorithm
6.7 Continuous Hidden State HMMs - LGSSMs
Просмотров 7853 года назад
6.7 Continuous Hidden State HMMs - LGSSMs
Wonderful lecture. Can't believe I'm finishing my PhD in meta-learning but only seeing this talk now !
The lecture is super clear, but I think it is a crime to use 6(1/2) for 6.5 😅
I have a question about the "normalizing constant". Suppose my friend's dad has a fever. I want to know what the probability of him having covid given that information. If I think it's more likely that he has covid, then my posterior is higher, and if I think the probability he would have a fever if he had covid, then my posterior is higher. But if I think the thermometer probably malfunctioned, then I should think it's less likely he has covid even with the fever measurement. But isn't that P(thermometer reading) the normalizing constant, and if it's lower such as in the case of a malfunction, then according to the formula, my posterior would be higher. So either the formula expresses my intuition incorrectly, or my understanding of the formula is not complete, the latter of which I intuitively believe is more likely. Can you help me understand this situation? Thank you!
Thank you; these are helpful and also different then what else is available... thanks again
Hi, this is a great course that I've stumbled upon that covers many important ML concepts from a probabilistic perspective. I am wondering whether there is any chance you could upload the sections 2-4 of your course?
Very glad that you are enjoying the lectures. I was one of two lecturers on the course and the other did not record their content I'm afraid. I hope to record them myself though at a later date.
@@ret2666 Cool, thanks for your reply!
Hello professor Turner, I really appreciate these lectures. They are one of the best on RUclips in my opinion. Could you please explain why you used the log prob instead of the prob? Thank you
Thanks for the message. We take logs in situations like this because: 1. Probabilities often involve a product of terms (here over all of the iid data points), and when we take logs this gets turned into a sum. Sums are easier to operate on e.g. differentiate or to implement on a computer (where a produce can lead to numerical underflow). 2. The distributions we work with typically are often in the exponential family (e.g. Gaussians) and taking logs leads to a simple form (for Gaussians we get quadratic functions back). Mathematically, taking logs does not change the optimum of the function as the log is a monotonic function, so the location of the optimum of a positive-valued function f(x) is in the same location as log f(x).
@@ret2666 Thanks for the quick reply and extensive elaboration!