Has anyone been able to make it converge? What hyperparameters did you modify?

Thanks great explanation, I wonder how the sampling with T can produce plausible images with just a single pass. I would expect to call sampling code recursively T times with timestamp 1 at each call and z is updated according to result of the previous call. Similar to autoregressive architectures

artificial neural network (ann) can be used to create fraud as well

you are a true hero

Amazing!

I have a question : why do you use only 2 times the gcn layer in 11:33 ?

There is a little error at 10:35, in the bottom of the fraction it's [Whi || Whk], not j. But super video, continue please!

is the resulting embedding for each time stamp or snapshot passed from layer to layer under the hood by the model?

You need to build a graph from first principles, not jut using libraries to transfom your data.

Sir how can we use it with NN?

Great video!

Thank you so much! I spent many days trying to understand the meaning behind GCN but couldn't get it. And amazingly, it only took me less than 3 minutes watching your example at 1:54 to grasp its meaning.

Thank you for the great video!

Very well prepared and carefully delivered, congratulations and many thanks for your help in understanding LIME.

Thanks for the video!!

It is magnificent video, after understand math and concept behind TSNE, then u can clear ur concept here thoroughly here in this video. 🌟🌟🌟🌟

As a robot myself, i can confirm that an image really is worth 16x16 words

An image is worth 16x16 words🗣🗣🗣🗣🗣🗣🗣💯💯💯💯💯💯💯🔥🔥🔥🔥🔥🔥🔥

Great job mate, keep it up

Awesome! Thanks for this video!

Could your uncertainty bands in the ensemble code be incorrect? In the plot_de function, you compute: stds = torch.stack(mus).std(axis=0).detach().numpy()**(1/2) Why is this there another sqrt there? shouldn't it just be: stds = torch.stack(mus).std(axis=0).detach().numpy()? The results are visually better when the sqrt is included, but to me it sounds mathematically incorrect, no?

Why are many straight lines in your videos not straight?

I love your way of explaining each topic, thank you! A series on AWS coming maybe? 😢

it is a crime that this doesn't get more views.

Excellent presentation. I have a question, can DICE be carried out in a model that has 3 classes instead of 2?

I loved the video, my friend. Huge explanation on the topic. I'll certainly watch the other videos in a near future. Thank you

Great video, really well explained! One thing I found suspicious is that in the forward diffusion process you generate noise like this "noise = torch.randn_like(x_0)". As far as I know, this samples the uniform distibution U(0,1) and not the standard Gaussian.

Why do you need to use weight if you have already oversampled the data?

Why do you say that x1+x2+x3 is a concatenation? Isn't it a element-wise addition?

Does anyone have problem accessing Colab?

Why is it "x" and not "self.x" ? And why self.training and not training?

Could you explain more in depth why we have 32 nodes and 9 features for this problem, please?

Amazing video! I have a question regarding one data point attributes. We have both x and y. The features themselves are the speeds which is the ground truth right? We want to train our model to predict into next 12 timestamps and compare it with values in x. So what is the significance of y (though in documentation, its given that y is the ground truth, both x and y should have the same values but they differ. Only difference that has to be is that x additionally has time of day as a second feature for each of the 12 timestamps per data point) Could you kindly clarify this?

thank you for your great explanation

You did not cover LDA and ICA

I found it hard to follow initially but after understanding GCNN thoroughly, this video is a gem.

I agree with your lecture and was very nice. How to apply for images like x-ray or ct scan

VonRueden Circle

You have put a lot of time into creating this videos and it is really valuable that after 3 years it is still very useful

Greta Video, thank you!

"... github in the link below....". Where is the link?

wow what a great video.Thank you, helped me a lot.

I just discovered your channel and THANK YOU!!

Adaline Pines

I really enjoy this video! It is so concise, comprehensive and beautiful! And thanks a lot for so many useful links for further learning.

So how are the edge features used in GCN algorithm? are they completely ignored? Because according to this explanation, only the node features take part in the convolution process

The best part of Vision transformers is inbuilt support interpretability as compared to CNN where we had to compute saliency maps.

Perfect video to understand GATs. However, I guess, you forgot to add sigmoid function when you demonstrate h1' as a sum of multiplications of hi* and attention values, in the last seconds of the video: 13:51

why are we adding time embedding to input features, like literally adding them together. Can a simple concatenation of input features and time embedding possible ? btw dope video thanks for sharing

I'm making a model from sx-stackoverflow of SNAP, it only has source target and timestamp, no clue how to make a dataset for TGN from that