I am working with neural networks for years now and every now and then I come back to the fundamentals to check my current understanding. Your videos are truly great! They are short, accurate, visually helpful and contain practical knowledge. Thank you very much for making these and I hope you get to make a full series!
Implementing padding in convolution would slow the whole algortihm down on architectures like the TPU. It also applies to most GPUs, you simply want to avoid branched code. It gives you huge performance benefits to prepare the data in memory and stream it through the computing system.
I love your visualization. I always end up drawing blocks when I try to understand how the data flows in an architecture, but your visualizations are to scale, have colors, and give a very good idea how the "volumes" of the data and parameters change throughout the model. It's not just an aid for understanding basic concepts like convolution - it's a bit of an overkill for that, but it's excellent for visualizing a large NNs. I wonder if it would be possible to create a "model explorer" that would allow you to explore interactively in 3D, like a video game. Maybe even with VR? But just 3D is probably plenty good.
"all the way back to convolutions origin in signal processing"? Looking quickly in the "history" section on wikipedia could tell you, that convolution was first invented by d'Alembert in 1754...
After the current series on convolution, my tentative plans are to make a series on attention, and I could make vision transformers part of that series. Thanks for the suggestion.
I am working with neural networks for years now and every now and then I come back to the fundamentals to check my current understanding. Your videos are truly great! They are short, accurate, visually helpful and contain practical knowledge. Thank you very much for making these and I hope you get to make a full series!
I just watched all your videos. Didn’t know anything about CNNs before, now I feel I have a solid foundation. Please make some more!
Underrated channel. You should have much more views.
Thank you so much for this video series. It really helped me with understanding convolution neural network.
Keep it up! These are amazing, straightforward technical videos that are invaluable for someone starting out.
That's a really good job, you have save me from struggling with the AI. Thank you from China
I just shared your channel with my AI professor and told her to include those videos within our course because it made everything clear! Thank you
Man your videos are revolutionary. Please do a set on transformers next, I have spent many hours trying to visualize them!
Implementing padding in convolution would slow the whole algortihm down on architectures like the TPU. It also applies to most GPUs, you simply want to avoid branched code. It gives you huge performance benefits to prepare the data in memory and stream it through the computing system.
I'm new to NN and these are really helpful 👍
Great explanation
Great explanation. Please make a video on 3d convolution and strides.
I love your visualization. I always end up drawing blocks when I try to understand how the data flows in an architecture, but your visualizations are to scale, have colors, and give a very good idea how the "volumes" of the data and parameters change throughout the model. It's not just an aid for understanding basic concepts like convolution - it's a bit of an overkill for that, but it's excellent for visualizing a large NNs. I wonder if it would be possible to create a "model explorer" that would allow you to explore interactively in 3D, like a video game. Maybe even with VR? But just 3D is probably plenty good.
Will you be explaining the training process ? ie how gradient decent works in CNN's
Also will you be explaining Max pooling?
"all the way back to convolutions origin in signal processing"? Looking quickly in the "history" section on wikipedia could tell you, that convolution was first invented by d'Alembert in 1754...
Amazing work! Could you also make videos about vision transformer? Thanks
After the current series on convolution, my tentative plans are to make a series on attention, and I could make vision transformers part of that series. Thanks for the suggestion.
@@animatedai Cool! Would love to see attention related animation tutorial as well!!
Can you upload one on normalization techniques in convolutional neural network ?
Or better on transformers in vision it is hard to get these concepts