Machine Learning / Deep Learning Tutorials for Programmers playlist: ua-cam.com/play/PLZbbT5o_s2xq7LwI2y8_QtvuXZedL6tQU.html Keras Machine Learning / Deep Learning Tutorial playlist: ua-cam.com/play/PLZbbT5o_s2xrwRnXk_yCPtnqqo4_u2YGL.html
I once heard a story about the military of some country who had developed a neural network to recognise enemy tanks. The model had really accurate test results, but once they deployed the model into a real situation it marked the friendly tanks as enemy tanks. It turned out that the country where it was developed used primarily pictures with a blue sky for the images of their own tanks while the images of enemy tanks were mostly taken on cloudy days. Coincidentally they tested the model in a real situation when it was cloudy. In other words, it never actually learned to recognise the friendly or enemy tanks, but instead it learned to recognise blue skies and cloudy skies. I thought of this story because of the model recognising left facing dogs while not recognising right facing dogs.
Thank you very much for this simple yet clear and very helpful video! I learnt some data augmentation methods and about considering which methods are reasonable to use!
Great series, just what I've been looking for, but my dog spends most of the day, on its back snoring on the floor. But seriously, loving these vids, likes, subscribed, and bell tapped.
{ "question": "Data augmentation can help with all of the following EXCEPT:", "choices": [ "Fine tuning batch size", "Effectively growing the data set", "Reducing overfitting", "Increasing data set diversity" ], "answer": "Fine tuning batch size", "creator": "Chris", "creationDate": "2019-12-13T03:56:31.512Z" }
On the example of horizontally flipping to have more left-facing dogs, how would this be done? If there are 2:1 right-facing to left-facing, and we horizontally 20% of them, we'd still expect 2:1 right-facing to left-facing (we would inadvertantly flip some left-facing as well). Would they need to be selected by hand?
Your video is very great and helpful. One question though, What if im creating a chatbot and then I used data augmentation technique for my "user inputs". For example: "How are you?" I data augmented it by using different variation of question like "how's your day?". Is that how data augmentation works for texts?
Another great video as always.. :) Which type of data are not suitable for data augmentation? If we take the LUNG CT images, which augmented methods will be useful to achieve more accurate trainig. Will flipflop works fine for this?
Thank you, Hiroshi :) In general, I would say that data augmentation can likely be applied in most scenarios. The only time that I can think where it would not be appropriate would be if the images are expected to be very uniform in presentation. For example, say that your images were portraits of people's faces all taken at the exact same angle with the exact zoom, lighting, background, etc. Then, say that your classification task at hand was to be able to predict on future images that were just like these with no variation in the things mentioned, like zoom, angle, etc. Then it would probably not be helpful to add data augmentation on the original data set since the data that you'll be later predicting on will not be varied in the ways that data augmentation would help your model be able to identify more easily. For lung CT images, yes I think horizontally flipping the images would be suitable. Maybe also zooming in/out. Something that would perhaps not be suitable would be vertically flipping the images. (Unless it's not uncommon practice to have lung CT images upside-down.)
@@deeplizard Mam you are our inspiration. How can we connect with you. We don't even know your name or anything. Can you please share your linkedin with us.
Machine Learning / Deep Learning Tutorials for Programmers playlist: ua-cam.com/play/PLZbbT5o_s2xq7LwI2y8_QtvuXZedL6tQU.html
Keras Machine Learning / Deep Learning Tutorial playlist: ua-cam.com/play/PLZbbT5o_s2xrwRnXk_yCPtnqqo4_u2YGL.html
I once heard a story about the military of some country who had developed a neural network to recognise enemy tanks. The model had really accurate test results, but once they deployed the model into a real situation it marked the friendly tanks as enemy tanks. It turned out that the country where it was developed used primarily pictures with a blue sky for the images of their own tanks while the images of enemy tanks were mostly taken on cloudy days. Coincidentally they tested the model in a real situation when it was cloudy. In other words, it never actually learned to recognise the friendly or enemy tanks, but instead it learned to recognise blue skies and cloudy skies.
I thought of this story because of the model recognising left facing dogs while not recognising right facing dogs.
Great example!
i like how you explain stuff so simply. thanks woman
I've never seen such positive reviews on any other lectures. You must be doing it right.
Thank you very much for this simple yet clear and very helpful video! I learnt some data augmentation methods and about considering which methods are reasonable to use!
you make everything so simple. thank u
Great series, just what I've been looking for, but my dog spends most of the day, on its back snoring on the floor.
But seriously, loving these vids, likes, subscribed, and bell tapped.
Thank you, Gibe! 😁
Great high level summaries, and a nice voice too :)
Thanks, David!
Hello from Peru, very nice videos
{
"question": "Data augmentation can help with all of the following EXCEPT:",
"choices": [
"Fine tuning batch size",
"Effectively growing the data set",
"Reducing overfitting",
"Increasing data set diversity"
],
"answer": "Fine tuning batch size",
"creator": "Chris",
"creationDate": "2019-12-13T03:56:31.512Z"
}
Thanks, Chris! Just added your question to deeplizard.com
Thanks for sharing your knowledge.
thank you very much for this clear and helpful explanation.
Excellent Intro. Thanks.
Simple and nice explanation, thanks!
On the example of horizontally flipping to have more left-facing dogs, how would this be done? If there are 2:1 right-facing to left-facing, and we horizontally 20% of them, we'd still expect 2:1 right-facing to left-facing (we would inadvertantly flip some left-facing as well). Would they need to be selected by hand?
Now I know the right term :) Thanks!
This is too good...thank you
hello I'm from Vietnam, thks for your videos , so nice !
You're welcome, Thuận! Vietnam is on my list for places I'd like to visit!
@@deeplizardthanks, hope you will have a great time here !
Your video is very great and helpful. One question though, What if im creating a chatbot and then I used data augmentation technique for my "user inputs". For example: "How are you?" I data augmented it by using different variation of question like "how's your day?". Is that how data augmentation works for texts?
Impressive
Thanks, from Pakistan.
thanks! very helpful
Is there any method,that data augmentated based in the existing excel file not other than images
Another great video as always.. :)
Which type of data are not suitable for data augmentation? If we take the LUNG CT images, which augmented methods will be useful to achieve more accurate trainig. Will flipflop works fine for this?
Thank you, Hiroshi :)
In general, I would say that data augmentation can likely be applied in most scenarios. The only time that I can think where it would not be appropriate would be if the images are expected to be very uniform in presentation.
For example, say that your images were portraits of people's faces all taken at the exact same angle with the exact zoom, lighting, background, etc. Then, say that your classification task at hand was to be able to predict on future images that were just like these with no variation in the things mentioned, like zoom, angle, etc. Then it would probably not be helpful to add data augmentation on the original data set since the data that you'll be later predicting on will not be varied in the ways that data augmentation would help your model be able to identify more easily.
For lung CT images, yes I think horizontally flipping the images would be suitable. Maybe also zooming in/out. Something that would perhaps not be suitable would be vertically flipping the images. (Unless it's not uncommon practice to have lung CT images upside-down.)
Thanks alot.... Very clear... I'm learning alot from you...
Thanks.
That was perfect
can you augment non-image data? say a time series?
mam can you please tell what video editing software you are using?
Hey Himanshu - I used Camtasia here.
en.wikipedia.org/wiki/Camtasia
@@deeplizard Mam you are our inspiration. How can we connect with you. We don't even know your name or anything. Can you please share your linkedin with us.
@@himanshupoddar1395 go and watch it on the description.
can we use supervised learning with data augmentation
yes
I mean with annotated images
Slide to the LEFT, slide to the RIGHT
Two hops this time 💃🎶
ty
i see what you tried to explain with the vertical flipping a dog example. but you obviously have no dog 😅.
Haha!