I swear this playlist is one of the best resources I have ever seen on these topics. Great explanation. Please continue to upload more of this great content. Much thanks for your time and outstanding effort.
Just wanted to say that these videos are really well done and the speaker really knows what she's talking about. Iam doing my PhD right now in mechanical engineering, using deep learning for modeling a production process (steel) and your videos really helped me to get a much better grips on what to tune and do with my model. Highly appreciated, thx a lot :)!
Very comprehensive and efficient survey of regularization. This brought together items I have seen in NVidia training and other training in a very organized fashion. I don't comment on YT videos often, but this one was worth it--well done.
damn this whole series is like a gold mine ... i was suspicious of how a so well know topic be covered in so less time ... the videos might be not good but happy to be proven the wrong. THESE ARE GOLD ... thank you @AssemblyAI & thank you very much Ma'am for helping.
Another absolutely fantastic, accessible teaching resource on a complex machine learning concept. I don't think there are any resources out there that can match the quality, accessibility and clarity this resource provides.
Wow so useful, thank you for the amazing content. Your can feel the confidence of the lecturer and her explanations are very clear. Watching all the playlist
Thank you for this explanation. Like many I’d imagine, I’ve bumped into these concepts predominately via my use of SD. It’s nice having an overview of what’s being conveyed so I can understand what’s happening without getting too bogged down in the minutiae.
Overfitting is frequently happening in my programs. I tried reducing the number of input parameters but I know it is not a good solution. I was familiar with L1 and L2 regularisation. This tutorial helped me better understanding of them and other common methods. I tried to decrease both the train and test errors but I was not successful by the use of regularisation. I hope to do it soon 🙂 Thanks for your illustrative explanations
I don't understand why we are multiplying inputs by the keep probability, are we not using the coefficients from training? Would they not be subsequently dropped in testing as well?
With L2 regularisation, do we add the weights^2 of the whole network to the final loss function, or while doing backprop and only the input weights of a node to the loss of that node?
You said that L1 regularisation encourages weights to be 0.0 and this could lead to not considering some of outputs, is this the same behavior of dropouts?
It is a similar behaviour to dropouts, yes! Both L1 and dropout can make a network sparse (not all neurons are connected to all other neurons). The way they achieve it is still different though.
I don't understand the purpose for regularization. The sole purpose of weights is to quantify the importance of a feature (strength of connection). So it's very much possible that one weight has much larger value than others because it's more important to the desired outcome in real life. But if you regularize. Then that weight loses its value and therefore might result in incorrect prediction.
That's the point of regularization. It is used when your model is overfitting as stated in the video. If the model is performing decent without regularizers then you probably shouldn't use regularizers as this could result in underfitting.
Variance is a mathematical term used in probability theory and stochastic for the measurement of data spreading around a mean. You are confusing your audience by abusing this term in a different manner when claiming that variance is the change rate of prediction per training data change. Don't do that !
I swear this playlist is one of the best resources I have ever seen on these topics. Great explanation. Please continue to upload more of this great content. Much thanks for your time and outstanding effort.
That's great! Glad to hear you liked it!
Just wanted to say that these videos are really well done and the speaker really knows what she's talking about. Iam doing my PhD right now in mechanical engineering, using deep learning for modeling a production process (steel) and your videos really helped me to get a much better grips on what to tune and do with my model. Highly appreciated, thx a lot :)!
You are very welcome Tim and thank you for the support! - Mısra
Very comprehensive and efficient survey of regularization. This brought together items I have seen in NVidia training and other training in a very organized fashion. I don't comment on YT videos often, but this one was worth it--well done.
damn this whole series is like a gold mine ... i was suspicious of how a so well know topic be covered in so less time ... the videos might be not good but happy to be proven the wrong. THESE ARE GOLD ... thank you @AssemblyAI & thank you very much Ma'am for helping.
Another absolutely fantastic, accessible teaching resource on a complex machine learning concept. I don't think there are any resources out there that can match the quality, accessibility and clarity this resource provides.
Wow so useful, thank you for the amazing content. Your can feel the confidence of the lecturer and her explanations are very clear. Watching all the playlist
Fantastic and very clear explanation- thank you !
Thank you for this explanation. Like many I’d imagine, I’ve bumped into these concepts predominately via my use of SD. It’s nice having an overview of what’s being conveyed so I can understand what’s happening without getting too bogged down in the minutiae.
Thanks a lot for the great explanation!
Overfitting is frequently happening in my programs. I tried reducing the number of input parameters but I know it is not a good solution. I was familiar with L1 and L2 regularisation. This tutorial helped me better understanding of them and other common methods. I tried to decrease both the train and test errors but I was not successful by the use of regularisation. I hope to do it soon 🙂 Thanks for your illustrative explanations
Your explanation is just amazing! ...
I'm a Research scholar from India, your videos are just awesome 👍
This is amazing series with concepts well explained. Lot of the other videos dwelve a lot on mathematical formulas without explaining the concepts.
Really to the point and excellently delivered.
Brief yet very clear and informative. Thank you.
You are very welcome Armin! - Mısra
Wow. Thank you so much for this!
Great job. The explanation is very clear and easy to understand.
Thank you Jacob!
Thank you! This gave a good intro b4 I started reading Ian Goodfellow.
Your videos are so good keep up the good work I have read and watched a lot of content explain it yours is the best
Great playlist, the contents are on the spot to each topic in a minimum time. Please keep outstanding work.🤘and thanks for the content.
You are very welcome! - Mısra
I liked this video very much. You explained all these techniques very well in my opinion.. Thank you..
Brief, Concise and Precise.
Thank you!
Very clear and precise explanation. Thanks :)
You're welcome :)
the best video with a clear explanation
Glad to hear it!
Absolutely clear explanation
Glad it was helpful!
This is very, very helpful. Great explanation Thank you.
Great content! Subscribed.
I had tears in my eyes. absolute gem of a video.
Thanks for the clear explanation.
So Good, SO Good, Oh My God! Thank you soooo much!
I love your teaching. Keep it up
It proves Beauty and Wisdom can coexist.
This playlist is treasure for me.
Awesome!
Thanks for this informative lecture
Amazing video. Very good content.
Wow, very clear
thanks you helped me
so nice and simple
Brilliant video
Thank you!
wow, thank you!!
Beauty with Brains 💐
Amazing! 😍😍
I really loved it
Awesome :)
Great content!
Thanks Ryan!
Awesome stuff from thy side...Danke shen....can you give the link to the playlist containing these lectures.....
Here it is: th-cam.com/video/dccdadl90vs/w-d-xo.html
You fantastic thanks
I don't understand why we are multiplying inputs by the keep probability, are we not using the coefficients from training? Would they not be subsequently dropped in testing as well?
With L2 regularisation, do we add the weights^2 of the whole network to the final loss function, or while doing backprop and only the input weights of a node to the loss of that node?
Great video! Also you sound like you are from Turkey. Am I correct?
Yes, that is correct :)
What if input features have multiple dimensions ie age and height, can we still use batch norm as first layer to normalize thr input data ?
Thanks.
You're welcome!
You said that L1 regularisation encourages weights to be 0.0 and this could lead to not considering some of outputs, is this the same behavior of dropouts?
It is a similar behaviour to dropouts, yes! Both L1 and dropout can make a network sparse (not all neurons are connected to all other neurons). The way they achieve it is still different though.
@@AssemblyAI thank you so much for your prompt answer
@@brahimmatougui1195 You are very welcome! -Mısra
Can I use this L1 regularization to overcome the maximized mutual information problem?
ty
you might have inclue batch size too
Great video. I’m currently making flash cards and this was a great resource.
Glad it was helpful!
great
I didn't understand why do we need 'keep probability '.
Forget about Regularisation,
I just came here to look thy beautiful lady❤❤❤
I don't understand the purpose for regularization. The sole purpose of weights is to quantify the importance of a feature (strength of connection). So it's very much possible that one weight has much larger value than others because it's more important to the desired outcome in real life. But if you regularize. Then that weight loses its value and therefore might result in incorrect prediction.
That's the point of regularization. It is used when your model is overfitting as stated in the video. If the model is performing decent without regularizers then you probably shouldn't use regularizers as this could result in underfitting.
I don't understand why don't we simply reduce no of layers and neurons in a neural network to get rid of OVERfitting.
That's just one of the ways. You can also try to reduce size of the model as you said or try data augmentation
Variance is a mathematical term used in probability theory and stochastic for the measurement of data spreading around a mean. You are confusing your audience by abusing this term in a different manner when claiming that variance is the change rate of prediction per training data change. Don't do that !
plz speak slowly