Відео

Not sure why this playlist doesn't come on top when one searches for ML on youtube. Andrew Ng might be a good researcher but he's not a very good teacher. Killian teaches in such a good manner, that I never once felt bored of felt as if I'm studying. Thanks Killian, You're a Gem.

Is there any way we can get access to the projects for this course?

Amazing!!!!!!!!!!!!!!!!!!!!!!!!

We have a quiz question in lecture notes : "How does k affect the classifier? What happens if k = n? What happens if k = 1?" I do not think it is discussed in lectures. In my opinion, k is the only hyperparameter in this algorithm. For k = n, we are taking mode of the entire dataset labels as the output for test point, where as for k =1 , it will be assigned label that of the closest nearest neighbor. I have a doubt here, as we are using distance metric, what if we have 2 points(for simplicity) that are at equal distance to test point and have different labels. What happens in that case for k = 1? Similarly, for k = n, if we have equal proportion of binary class labels, how does mode works in that case?

thank you professor .

Everyone seems to have a different version. AdaGrad doesn't always work. The sum of the dot product of the gradient gets too big UNLESS one scales it down. Also, AdaGrad works best with a line search. All variations work best with a line search.

i am wondering teenager hahaha!

RIP harambe

wauuw in my second video already i scrolled through the videos it seems it will be much better to learn this parallel with the pattern recognition book of bishop

He is Hermann Minkowski and was Einstein's teacher. Minkowski metric is the metric of flat space time and forms the backbone of special relativity. The ideas developed by Minkowski were later extended by Einstein to develop the theory of general relativity.

OMG! I'm in love with this guy! the energy he has makes me want to learn ML!

the Donald Trump bits were very funny!

Thank you for this conference, I learned a lot! I do have one question though. In the course, it is shown how to diagnose the ML model to balance the bias/variance trade-off, but what about noise? How is it possible to know if the error of the model comes from significant noise in the dataset?

Alrighht... hello everybody...

You are a J word.

I've never encountered anything better than these playlist. Thank you, Professor, for the detailed explanation and, most importantly, for presenting it in such an engaging way that it sparks a deep passion in everyone. Here I am in 2024, following since lecture one and feeling how much I have developed after your lectures.

What was the question at 14:30 anyone know? Brilliant lecture - easily a new all time favourite.

Great lecture, thanks Professor! Integrating the online Certificate with deeper math and concepts

He's such a good lecturer that 2/3rds of students show up in person.

A minor suggestion with the charts showing the distance vs dimensions at around the 29:00 minute mark.... I think at least one student was a bit misled because they didn't notice the x-axis changes so drastically because the peak of distribution visually appears to be in the same spot on each of the 6 charts.

Supervised learning - making predictions from data - dataset D with n data points and their outputs; the feature vector and label - the data points are a sample from some distribution - the aim is to get a function from the dataset that maps x to y Examples of y (labelling) - binary classification: there or not label - k classes: multiple labels -

Loved that "the answer will always be Gaussian, the whole lecture!" moment.

Hi Professor Weinberger, it looks like this lecture is about a smart algorithm created by smart people which can classify datas into two classes. But in the first introduction lecture you mentioned that machine learning is about computer learning to design a program by itself to achieve our goal. So I' confused what's the relationship between this perceptron hyperplane algorithm with machine learning? It looks like we human just design this algorithm and code it into a program and feed it into a computer to solve the classification problem...

Your English is so perfect.

Hi Professor, thank you so much for the lecture. I wonder if it's possible that AdaBoost stops when training error is zero? Because I see from your demo, after training error is close to zero and exponential loss goes smaller and smaller, the test error doesn't change too much. I guess we don't need to waste time on letting exponential loss smaller and smaller.

16:40

17:54 volunteers 🤣 Thanks prof for the fun and interesting lecture, got to revise these fundamentals quickly 🙏

thank you for these lectures!

Hello, Thank you for the lecture. Why is the variance equal for all points? 17:23 Is this an assumption that we are taking?

Best vedio on Bias-Variance Decomposition ❤

I was looking for an answer that was quite technical in another video but I got hooked. Thank you so much for providing such great knowledge.

the party example and the demo makes the algorithm so easy that a mid school student can understand to

*says some stuff* "aaaAAAHHhh" *says some more stuff* "aaAAAHHhhhh"

Hi All, can someone give some more information about the notation at 31:40? What does the indicator “I” mean? Thanks

Very beautiful algorithm

Wish we can get a new lecture series

I just want to say the way you teach Killian, I haven't seen anyone this talented who can explain a concept in such a clear manner. Even people with minimal background get the big picture of what you're trying to say. Please keep uploading these excellent lectures

Hi Professor. Thank you for sharing the video. I am now using Gaussian Process Regression in physics field. One thing I noticed is that even though there exists specific loss function for GPR, many people use root-mean-squared-error as loss function. Is there any rule to choose loss function and regularization?

The intuition here was so so satisfying. The way it all comes together at the end, when he points out that the sigmoidal functions people used to use (because of emulating neuronal activation functions) have these flat parts which slow down the gradient. Not only is the slowed learning bad, but that slowed learning dampens the ability of the noisy SGD to escape the thin deep wells which represent ideal parameters only for a SPECIFIC data set. I.e. the thin deep wells = overfitting, the noise of SGD escapes them along with big alpha, and a slowed gradient from the sigmoidal flat parts causes an effective reduction in learning rate, which leads to getting trapped in the wells even with SGD, which causes overfitting. Just awesome.

@kilianweinberger698 Sir, How do we ensure that adding pos encoding does not distort the original embedding too much? or how is that the sums of embedding and positional encoding of different tokens do not collide?

For non-parametric regression, when we test a data, we always need the whole training dataset right? Does it mean that we need to use a lot of memory to save the model, i.e. the training dataset?

i have a question how we know the best sequence of features that we should use in each depth layer because if we want to try each one and optimize with 30 to 40 features will take forever , or how we can do this for m features because i can really visual how this work.

I was thinking about modeling the prediciton of y given x with a Gaussian. Are these observations/reasoning steps correct? I understand the Gaussianess comes because we have a true linear function that perfectly models the relationship between X and Y, but it is uknown to us. But we have data (D) that we assume comes from sampling the true distribution (P). Now, we only have this limited sample of data, so it's reasonable to model the noise as Gaussian. This means that for a given x, our prediction y actually belongs to a Gaussian distribution, but since we only have this "single" sample D of the true data distribution, our best bet is to assign this y as the expectation of the true Gaussian. Which results in us predicting y as the final prediction (also because a good estimator of the expectation is the average, I guess). Now, I have explained how in the end we are going to fit the model to the data and predict that, so why do we have to model the noise in the model? Why not make it purely an optimization problem? I guess more like the DL approach.

hi nice

I have a question regarding masked multi-head attention around 53:30, if the outputs are generated one by one, then how can the word 'bananas' knows about the word 'cherries' (because at the time bananas is generated, cherries is not yet generated) and be modified by it? ie., why do we have to worry about cherries modifying bananas (aka having information about the future) if cherries hasn't even existed at that point?

I raise my hand. Why you assume any type of distribution when discussing? What if I don't know that formula? But what I see is nH and nT. Why not work with those?

legendary!

can someone tell where's the lecture in which he proved K nearest algorithm which he mentioned @5:09

Hi if anyone can advice, 40:00, so since we have close form solution, if we need implement ridge regression with kernel, we don't use Gradient Descent, but the close form directly, is that correct? because at lecture 21 at 16:40 prof. Kilian gave the recursive gradient descent solution, and I tried implement it which diverges very quickly, sensitive to the step size. Thank you!

Can't thank you enough for these lectures Professor Weinberger!